This paper has been accepted for publication at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, 2021. ©IEEE Time Lens: Event-based Video Frame Interpolation Stepan Tulyakov*;1Daniel Gehrig;2Stamatios Georgoulis1Julius Erbach1 Mathias Gehrig2Yuanyou Li1Davide Scaramuzza2 1Huawei Technologies, Zurich Research Center 2Dept. of Informatics, Univ. of Zurich and Dept. of Neuroinformatics, Univ. of Zurich and ETH Zurich Figure 1: Qualitative results comparing our proposed method, Time Lens, with DAIN [3] and BMBC [29]. Our method can interpolate frames in highly-dynamic scenes, such as while spinning an umbrella (top row) and bursting a balloon (bottom row). It does this by combining events (b) and frames (a). Abstract State-of-the-art frame interpolation methods generate intermediate frames by inferring object motions in the image from consecutive key-frames. In the absence of additional information, first-order approximations, i.e. optical flow, must be used, but this choice restricts the types of motions that can be modeled, leading to errors in highly dynamic scenarios. Event cameras are novel sensors that address this limitation by providing auxiliary visual information in the blind-time between frames. They asynchronously measure per-pixel brightness changes and do this with high temporal resolution and low latency. Event-based frame interpolation methods typically adopt a synthesis-based approach, where predicted frame residuals are directly applied to the key-frames. However, while these approaches can capture non-linear motions they suffer from ghosting and perform poorly in low-texture regions with few events. Thus, synthesis-based and flow-based approaches are complementary. In this work, we introduce *indicates equal contributionTime Lens , a novel method that leverages the advantages of both. We extensively evaluate our method on three synthetic and two real benchmarks where we show an up to 5.21 dB improvement in terms of PSNR over state-of-the-art frame-based and event-based methods. Finally, we release a new large-scale dataset in highly dynamic scenarios, aimed at pushing the limits of existing methods. Multimedia Material The High-Speed Event and RGB (HS-ERGB) dataset and evaluation code can be found at: http://rpg.ifi. uzh.ch/timelens 1. Introduction Many things in real life can happen in the blink of an eye. A hummingbird flapping its wings, a cheetah accel- erating towards its prey, a tricky stunt with the skateboard, or even a baby taking its first steps. Capturing these mo- ments as high-resolution videos with high frame rates typi-arXiv:2106.07286v1 [cs.CV] 14 Jun 2021cally requires professional high-speed cameras, that are in- accessible to casual users. Modern mobile device producers have tried to incorporate more affordable sensors with sim- ilar functionalities into their systems, but they still suffer from the large memory requirements and high power con- sumption associated with these sensors. Video Frame Interpolation (VFI) addresses this problem, by converting videos with moderate frame rates high frame rate videos in post-processing. In theory, any number of new frames can be generated between two keyframes of the input video. Therefore, VFI is an important problem in video processing with many applications, ranging from super slow motion [10] to video compression [42]. Frame-based interpolation approaches relying solely on input from a conventional frame-based camera that records frames synchronously and at a fixed rate. There are several classes of such methods that we describe below. Warping-based approaches [20, 10, 44, 21, 29] combine optical flow estimation [8, 16, 36] with image warping [9], to generate intermediate frames in-between two consecutive key frames. More specifically, under the assumptions of lin- ear motion and brightness constancy between frames, these works compute optical flow and warp the input keyframe(s) to the target frame, while leveraging concepts, like contex- tual information [20], visibility maps [10], spatial trans- former networks [44], forward warping [21], or dynamic blending filters [29], to improve the results. While most of these approaches assume linear motion, some recent works assume quadratic [43] or cubic [5] motions. Although these methods can address non-linear motions, they are still lim- ited by their order, failing to capture arbitrary motion. Kernel-based approaches [22, 23] avoid the explicit mo- tion estimation and warping stages of warping-based ap- proaches. Instead, they model VFI as local convolution over the input keyframes, where the convolutional kernel is esti- mated from the keyframes. This approach is more robust to motion blur and light changes. Alternatively, phase-based approaches [18] pose VFI as a phase shift estimation prob- lem, where a neural network decoder directly estimates the phase decomposition of the intermediate frame. However, while these methods can in theory model arbitrary motion, in practice they do not scale to large motions due to the lo- cality of the convolution kernels. In general, all frame-based approaches assume simplis- tic motion models (e.g. linear) due to the absence of vi- sual information in the blind-time between frames, which poses a fundamental limitation of purely frame-based VFI approaches. In particular, the simplifying assumptions rely on brightness and appearance constancy between frames, which limits their applicability in highly dynamic scenar- ios such as ( i) for non-linear motions between the input keyframes, ( ii) when there are changes in illumination or motion blur, and ( iii) non-rigid motions and new objects ap-pearing in the scene between keyframes. Multi-camera approaches. To overcome this limita- tion, some works seek to combine inputs from several frame-based cameras with different spatio-temporal trade- offs. For example, [1] combined low-resolution video with high resolution still images, whereas [25] fused a low- resolution high frame rate video with a high resolution low frame rate video. Both approaches can recover the miss- ing visual information necessary to reconstruct true object motions, but this comes at the cost of a bulkier form factor, higher power consumption, and a larger memory footprint. Event-based approaches. Compared to standard frame- based cameras, event cameras [14, 4] do not incur the afore- mentioned costs. They are novel sensors that only report the per-pixel intensity changes, as opposed to the full intensity images and do this with high temporal resolution and low latency on the order of microseconds. The resulting output is an asynchronous stream of binary “events” which can be considered a compressed representation of the true visual signal. These properties render them useful for VFI under highly dynamic scenarios (e.g. high-speed non-linear mo- tion, or challenging illumination). Events-only approaches reconstruct high frame rate videos directly from the stream of incoming events using GANs [38], RNNs [32, 33, 34], or even self-supervised CNNs [28], and can be thought of as a proxy to the VFI task. However, since the integration of intensity gradients into an intensity frame is an ill-posed problem, the global contrast of the interpolated frames is usually miscalculated. Moreover, as in event cameras intensity edges are only ex- posed when they move, the interpolation results are also de- pendent on the motion. Events-plus-frames approaches. As certain event cam- eras such as the Dynamic and Active VIsion Sensor (DA VIS) [4] can simultaneously output the event stream and intensity images – the latter at low frame rates and prone to the same issues as frame-based cameras (e.g. motion blur) – several works [26, 41, 11, 37] use both streams of information. Typically, these works tackle VFI in conjunc- tion with de-blurring, de-noising, super-resolution, or other relevant tasks. They synthesize intermediate frames by accumulating temporal brightness changes, represented by events, from the input keyframes and applying them to the key frames. While these methods can handle illumination changes and non-linear motion they still perform poorly compared to the frame-based methods (please see § 3.2), as due to the inherent instability of the contrast threshold and sensor noise, not all brightness changes are accurately registered as events. Our contributions are as follows 1. We address the limitations of all aforementioned methods by introducing a CNN framework, named Time Lens , that marries the advantages of warping-Figure 2: Proposed event-based VFI approach. and synthesis-based interpolation approaches. In our framework, we use a synthesis-based approach to ground and refine results of high-quality warping- based approach and provide the ability to handle illu- mination changes and new objects appearing between keyframes (refer Fig. 7), 2. We introduce a new warping-based interpolation ap- proach that estimates motion from events, rather than frames and thus has several advantages: it is more ro- bust to motion blur and can estimate non-linear mo- tion between frames. Moreover, the proposed method provides a higher quality interpolation compared to synthesis-based methods that use events when event information is not sufficient or noisy. 3. We empirically show that the proposed Time Lens greatly outperforms state-of-the-art frame-based and event-based methods, published over recent months, on three synthetic and two real benchmarks where we show an up to 5.21 dB improvement in terms of PSNR. 2. Method Problem formulation. Let us assume an event-based VFI setting, where we are given as input the left I0and rightI1RGB key frames, as well as the left E0!and right E!1event sequences , and we aim to interpolate (one or more) new frames ^Iat random timesteps in-between the key frames. Note that, the event sequences ( E0!,E!1) contain all asynchronous events that are triggered from the moment the respective (left I0or rightI1) key RGB frame is synchronously sampled, till the timestep at which we want to interpolate a new frame ^I. Fig. 2 illustrates the proposed event-based VFI setting. System overview. To tackle the problem under consid- eration we propose a learning-based framework, namely Time Lens , that consists of four dedicated modules that serve complementary interpolation schemes, i.e. warping- based and synthesis-based interpolation. In particular, (1) thewarping-based interpolation module estimates a new frame by warping the boundary RGB keyframes using op- tical flow estimated from the respective event sequence; (2) thewarping refinement module aims to improve this esti- mate by computing residual flow; (3) the interpolation by synthesis module estimates a new frame by directly fusing the input information from the boundary keyframes and the event sequences; finally (4) the attention-based averaging module aims to optimally combine the warping-based andsynthesis-based results. In doing so, Time Lens marries the advantages of warping- and synthesis-based interpola- tion techniques, allowing us to generate new frames with color and high textural details while handling non-linear motion, light changes, and motion blur. The workflow of our method is shown in Fig. 3a. All modules of the proposed method use the same back- bone architecture, which is an hourglass network with skip connections between the contracting and expanding parts, similar to [10]. The backbone architecture is described in more detail in the supplementary materials. Regarding the learning representation [7] used to encode the event sequences, all modules use the voxel grid representation. Specifically, for event sequence E0!endwe compute a voxel gridV0!endfollowing the procedure described in [46]. In the following paragraphs, we analyze each mod- ule and its scope within the overall framework. Interpolation by synthesis , as shown in Fig. 3b, directly regresses a new frame ^Isyngiven the left I0and rightI1 RGB keyframes and events sequences E0!andE!1re- spectively. The merits of this interpolation scheme lie in its ability to handle changes in lighting, such as water re- flections in Fig. 6 and a sudden appearance of new objects in the scene, because unlike warping-based method, it does not rely on the brightness constancy assumption. Its main drawback is the distortion of image edges and textures when event information is noisy or insufficient because of high contrast thresholds, e.g. triggered by the book in Fig. 6. Warping-based interpolation , shown in Fig. 3d, first estimates the optical flow F!0andF!1between a la- tent new frame ^Iand boundary keyframes I0andI1using eventsE!0andE!1respectively. We compute E!0, by reversing the event sequence E0!, as shown in Fig. 4. Then our method uses computed optical flow to warp the boundary keyframes in timestep using differentiable in- terpolation [9], which in turn produces two new frame esti- mates ^Iwarp 0!and^Iwarp 1!. The major difference of our approach from the tradi- tional warping-based interpolation methods [20, 10, 21, 43], is that the latter compute optical flow between keyframes using the frames themselves and then approximate opti- cal flow between the latent middle frame and boundary by using a linear motion assumption. This approach does not work when motion between frames is non-linear and keyframes suffer from motion blur. By contrast, our ap- proach computes the optical flow from the events, and thus can naturally handle blur and non-linear motion. Although events are sparse, the resulting flow is sufficiently dense as shown in Fig. 3d, especially in textured areas with dominant mostion, which is most important for interpolation. Moreover, the warping-based interpolation approach re- lying on events also works better than synthesis-based method in the scenarios when event data is noisy or not(a) Overview of the proposed method. (b) Interpolation by synthesis module. (c) Attention-based averaging module. (d) Warping-based interpolation module. (e) Warping refinement module. Figure 3: Structure of the proposed method. The overall workflow of the method is shown in Fig. 3a and individual modules are shown in Fig. 3d, 3b, 3e and 3c. In the figures we also show loss function that we use to train each module. We show similar modules in the same color across the figures. Figure 4: Example of an event sequence reversal. sufficient due to high contrast thresholds, e.g. the book in Fig. 6. On the down side, this method still relies on the brightness constancy assumption for optical flow estimation and thus can not handle brightness changes and new ob- jects appearing between keyframes, e.g. water reflections in Fig. 6. Warping refinement module computes refined interpo- lated frames, ^Irefine 0!and^Irefine 1!, by estimating residual op- tical flow, F!0andF!1respectively, between the warping-based interpolation results, ^Iwarp 0!and^Iwarp 1!, and the synthesis result ^Isyn . It then proceeds by warping ^Iwarp 0! and^Iwarp 1!for a second time using the estimated residual optical flow, as shown in Fig. 3e. The refinement module draws inspiration from the success of optical flow and dis- parity refinement modules in [8, 27], and also by our ob- servation that the synthesis interpolation results are usually perfectly aligned with the ground-truth new frame. Besides computing residual flow, the warping refinement module also performs inpainting of the occluded areas, by fillingthem with values from nearby regions. Finally, the attention averaging module, shown in Fig. 3c, blends in a pixel-wise manner the results of synthe- sis^Isyn and warping-based interpolation ^Irefine 0!and^Irefine 1!to achieve final interpolation result ^I. This module leverages the complementarity of the warping- and synthesis-based interpolation methods and produces a final result, which is better than the results of both methods by 1.73 dB in PSNR as shown in Tab. 1 and illustrated in Fig. 6. A similar strategy was used in [21, 10], however these works only blended the warping-based interpolation results to fill the occluded regions, while we blend both warping and synthesis-based results, and thus can also handle light changes. We estimate the blending coefficients using an at- tention network that takes as an input the interpolation re- sults, ^Irefine 0!,^Irefine 1!and^Isyn, the optical flow results F!0 andF!1and bi-linear coefficient , that depends on the position of the new frame as a channel with constant value. 2.1. High Speed Events-RGB (HS-ERGB) dataset Due to the lack of available datasets that combine synchronized, high-resolution event cameras and standard RGB cameras, we build a hardware synchronized hybrid sensor which combines a high-resolution event camera withEvent Camera Prophesee Gen4M 720p Resolution 1280 720 RGB Camera FLIR BlackFly S Resolution: 1440 1080 2.5 cm baselineFigure 5: Illustration of the dual camera setup. It comprises a Prophesee Gen4 720p monochrome event camera (top) and a FLIR BlackFly S RGB camera (bottom). Both cam- eras are hardware synchronized with a baseline of 2:5 cm . a high resolution and high-speed color camera. We use this hybrid sensor to record a new large-scale dataset which we term the High-Speed Events and RGB (HS-ERGB) dataset which we use to validate our video frame interpolation ap- proach. The hybrid camera setup is illustrated in Fig. 5. It features a Prophesee Gen4 (1280 720) event camera (Fig. 5 top) and a FLIR BlackFly S global shutter RGB cam- era (1440 1080) (Fig. 5 bottom), separated by a baseline of2:5 cm . Both cameras are hardware synchronized and share a similar field of view (FoV). We provide a detailed comparison of our setup against the commercially available DA VIS 346 [4] and the recently introduced setup [40] in the appendix.Compared to both [4] and [40] our setup is able to record events at much higher resolution (1280 720 vs. 240 180 or 346 260) and standard frames at much higher framerate (225 FPS vs. 40 FPS or 35 FPS) and with a higher dynamic range (71.45 dB vs. 55 dB or 60 dB). Moreover, standard frames have a higher resolution com- pared to the DA VIS sensor (1440 1080 vs. 240 180) and provide color. The higher dynamic range and frame rate, enable us to more accurately compare event cameras with standard cameras in highly dynamic scenarios and high dy- namic range. Both cameras are hardware synchronized and aligned via rectification and global alignment. For more synchronization and alignment details see the appendix. We record data in a variety of conditions, both indoors and outdoors. Sequences were recorded outdoors with ex- posure times as low as 100µsor indoors with exposure times up to 1000 µs. The dataset features frame rates of 160 FPS, which is much higher than previous datasets, en- abling larger frame skips with ground truth color frames. The dataset includes highly dynamic close scenes with non- linear motions and far-away scenes featuring mainly cam- era ego-motion. For far-away scenes, stereo rectification is sufficient for good per-pixel alignment. For each sequence, alignment is performed depending on the depth either by stereo rectification or using feature-based homography esti- mation.To this end, we perform standard stereo calibration between RGB images and E2VID [32] reconstructions and rectify the images and events accordingly. For the dynamic close scenes, we additionally estimate a global homogra- phy by matching SIFT features [17] between these two im-ages. Note that for feature-based alignment to work well, the camera must be static and objects of interest should only move in a fronto-parallel plane at a predetermined depth. While recording we made sure to follow these constraints. For a more detailed dataset overview we refer to the sup- plementary material. 3. Experiments All experiments in this work are done using the Py- Torch framework [30]. For training, we use the Adam optimizer[12] with standard settings, batches of size 4 and learning rate 104, which we decrease by a factor of 10ev- ery 12 epoch. We train each module for 27 epoch. For the training, we use large dataset with synthetic events gener- ated from Vimeo90k septuplet dataset [44] using the video to events method [6], based on the event simulator from [31]. We train the network by adding and training modules one by one, while freezing the weights of all previously trained modules. We train modules in the following or- der: synthesis-based interpolation, warping-based interpo- lation, warping refinement, and attention averaging mod- ules. We adopted this training because end-to-end training from scratch does not converge, and fine-tuning of the en- tire network after pretraining only marginally improved the results. We supervise our network with perceptual [45] and L1losses as shown in Fig. 3b, 3d, 3e and 3c. We fine-tune our network on real data module-by-module in the order of training. To measure the quality of interpolated images we use structural similarity (SSIM) [39] and peak signal to noise ratio (PSNR) metrics. Note, that the computational complexity of our interpo- lation method is among the best: on our machine for image resolutions of 640480, a single interpolation on the GPU takes 878 ms for DAIN [3], 404 ms for BMBC [29], 138 ms for ours, 84 ms for RRIN [13], 73 ms for Super SloMo [10] and 33 ms for LEDVDI [15] methods. 3.1. Ablation study To study the contribution of every module of the pro- posed method to the final interpolation, we investigate the interpolation quality after each module in Fig. 3a, and re- port their results in Tab. 1. The table shows two notable re- sults. First, it shows that adding a warping refinement block after the simple warping block significantly improves the interpolation result. Second, it shows that by attention aver- aging synthesis-based and warping-based results, the inter- polations are improved by 1.7 dB in terms of PSNR. This is because the attention averaging module combines the ad- vantages of both methods. To highlight this further, we il- lustrate example reconstructions from these two modules in Fig. 6. As can be seen, the warping-based module excels at reconstructing textures in non-occluded areas (fourth col- umn) while the synthesis module performs better in regionswith difficult lighting conditions (fifth column). The atten- tion module successfully combines the best parts of both modules (first column). Figure 6: Complementarity of warping- and synthesis- based interpolation. Table 1: Quality of interpolation after each module on Vimeo90k (denoising) validation set. For SSIM and PSNR we show mean and one standard deviation. The best result is highlighted. Module PSNR SSIM Warping interpolation 26.68 3.68 0.926 0.041 Interpolation by synthesis 34.10 3.98 0.964 0.029 Warping refinement 33.02 3.76 0.963 0.026 Attention averaging (ours) 35.83 3.70 0.976 0.019 3.2. Benchmarking Synthetic datasets. We compare the proposed method, which we call Time Lens , to four state-of-the-art frame-based interpolation methods DAIN [3],RRIN [13], BMBC [29], SuperSloMo [10], event-based video recon- struction method E2VID [33] and two event and frame- based methods EDI [26] and LEDVDI [15] on pop- ular video interpolation benchmark datasets, such as Vimeo90k (interpolation) [44], Middlebury [2]. During the evaluation, we take original video sequence, skip 1 or 3 frames respectively, reconstruct them using interpola- tion method and compare to ground truth skipped frames. Events for event-based methods we simulate using [6] from the skipped frames. We do not fine-tune the methods for each dataset but simply use pre-trained models provided by the authors. We summarise the results in Tab. 2. As we can see, the proposed method outperforms other method across datasets in terms of average PSNR (up to 8.82 dB improvement) and SSIM scores (up to 0.192 im- provement). As before these improvements stem from the use of auxiliary events during the prediction stage which allow our method to perform accurate frame interpolation, event for very large non-linear motions. Also, it has signif- icantly lower standard deviation of the PSNR (2.53 dB vs. 4.96 dB) and SSIM (0.025 vs. 0.112) scores, which sug- gests more consistent performance across examples. Also,we can see that PSNR and SSIM scores of the proposed method degrades to much lesser degree than scores of the frame-based methods (up to 1.6 dB vs. up to 5.4 dB), as we skip and attempt to reconstruct more frames. This suggests that our method is more robust to non-linear motion than frame-based methods. High Quality Frames (HQF) dataset. We also evalu- ate our method on High Quality Frames (HQF) dataset [35] collected using DA VIS240 event camera that consists of video sequences without blur and saturation. During eval- uation, we use the same methodology as for the synthetic datasets, with the only difference that in this case we use real events. In the evaluation, we consider two versions of our method: Time Lens-syn , which we trained only on syn- thetic data, and Time Lens-real , which we trained on syn- thetic data and fine-tuned on real event data from our own DA VIS346 camera. We summarise our results in Tab. 3. The results on the dataset are consistent with the re- sults on the synthetic datasets: the proposed method outper- forms state-of-the-art frame-based methods and produces more consistent results over examples. As we increase the number of frames that we skip, the performance gap be- tween the proposed method and the other methods widens from 2.53 dB to 4.25 dB, also the results of other methods become less consistent which is reflected in higher devia- tion of PSNR and SSIM scores. For a more detailed dis- cussion about the impact of frame skip length and perfor- mance, see the appendix. Interestingly, fine-tuning of the proposed method on real event data, captured by another camera, greatly boosts the performance of our method by an average of 1.94 dB. This suggest that existence of large domain gap between synthetic and real event data. High Speed Event-RGB dataset. Finally, we evaluate our method on our dataset introduced in § 2.1. As clear from Tab. 4, our method, again significantly outperforms frame- based and frame-plus-event-based competitors. In Fig. 7 we show several examples from the HS-ERGB test set which show that, compared to competing frame-based method, our method can interpolate frames in the case of nonlin- ear (“Umbrella” sequence) and non-rigid motion (“Water Bomb”), and also handle illumination changes (“Fountain Schaffhauserplatz” and “Fountain Bellevue”). 4. Conclusion In this work, we introduce Time Lens, a method that can show us what happens in the blind-time between two intensity frames using high temporal resolution in- formation from an event camera. It works by leveraging the advantages of synthesis-based approaches, which can handle changing illumination conditions and non-rigid motions, and flow-based approach, relying on motion estimation from events. It is therefore robust to motion blur and non-linear motions. The proposed method achievesTable 2: Results on standard video interpolation benchmarks such as Middlebury [2],Vimeo90k (interpolation) [44] and GoPro [19]. In all cases, we use a test subset of the datasets. To compute SSIM and PSNR, we downsample the original video and reconstruct the skipped frames. For Middlebury and Vimeo90k (interpolation), we skip 1 and 3 frames, and for GoPro we skip 7 and 15 frames due its its high frame rate of 240 FPS. Uses frames andUses events indicate if a method uses frames and events for interpolation. For event-based methods we generate events from the skipped frames using the event simulator [6]. Color indicates if a method works with color frames. For SSIM and PSNR we show mean and one standard deviation. Note, that we can not produce results with 3 skips on the Vimeo90k dataset, since it consists of frame triplet. We show the best result in each column in bold and the second-best using underscore text. Method Uses frames Uses events Color PSNR SSIM PSNR SSIM Middlebury [2] 1 frame skip 3 frames skips DAIN [3] 4 8 4 30.875.38 0.899 0.110 26.674.53 0.838 0.130 SuperSloMo [10] 4 8 4 29.755.35 0.880 0.112 26.43 5.30 0.823 0.141 RRIN [13] 4 8 4 31.085.55 0.8960.112 27.18 5.57 0.8370.142 BMBC [29] 4 8 4 30.836.01 0.897 0.111 26.86 5.82 0.834 0.144 E2VID [32] 8 4 8 11.262.82 0.427 0.184 26.86 5.82 0.834 0.144 EDI [26] 4 4 8 19.722.95 0.725 0.155 18.44 2.52 0.669 0.173 Time Lens (ours) 4 4 4 33.273.11 0.929 0.027 32.13 2.81 0.908 0.039 Vimeo90k (interpolation) [44] 1 frame skip 3 frames skips DAIN [3] 4 8 4 34.204.43 0.962 0.023 - - SuperSloMo [10] 4 8 4 32.934.23 0.948 0.035 - - RRIN [13] 4 8 4 34.724.40 0.9620.029 - - BMBC [29] 4 8 4 34.564.40 0.962 0.024 - - E2VID [32] 8 4 8 10.082.89 0.395 0.141 - - EDI [26] 4 4 8 20.743.31 0.748 0.140 - - Time Lens (ours) 4 4 4 36.313.11 0.962 0.024 - - GoPro [19] 7 frames skip 15 frames skips DAIN [3] 4 8 4 28.814.20 0.876 0.117 24.39 4.69 0.7360.173 SuperSloMo [10] 4 8 4 28.984.30 0.875 0.118 24.38 4.78 0.747 0.177 RRIN [13] 4 8 4 28.964.38 0.876 0.119 24.324.80 0.749 0.175 BMBC [29] 4 8 4 29.084.58 0.8750.120 23.68 4.69 0.736 0.174 E2VID [32] 8 4 8 9.742.11 0.549 0.094 9.75 2.11 0.549 0.094 EDI [26] 4 4 8 18.792.03 0.670 0.144 17.45 2.23 0.603 0.149 Time Lens (ours) 4 4 4 34.811.63 0.959 0.012 33.21 2.00 0.942 0.023 Table 3: Benchmarking on the High Quality Frames (HQF) DA VIS240 dataset. We do not fine-tune our method and other methods and use models provided by the authors. We evaluate methods on all sequences of the dataset. To compute SSIM and PSNR, we downsample the original video by skip 1 and 3 frames, reconstruct these frames and compare them to the skipped frames. In Uses frames andUses events columns we specify if a method uses frames and events for interpolation. In the Color column, we indicate if a method works with color frames. In the table, we present two versions of our method: Time Lens-syn , which we trained only on synthetic data, and Time Lens-real , which we trained on synthetic data and fine- tuned on real event data from our own DA VIS346 camera. For SSIM and PSNR, we show mean and one standard deviation. We show the best result in each column in bold and the second-best using underscore text. Method Uses frames Uses events Color PSNR SSIM PSNR SSIM 1 frame skip 3 frames skips DAIN [3] 4 8 4 29.826.91 0.875 0.124 26.10 7.52 0.782 0.185 SuperSloMo [10] 4 8 4 28.766.13 0.861 0.132 25.54 7.13 0.761 0.204 RRIN [13] 4 8 4 29.767.15 0.874 0.132 26.11 7.84 0.778 0.200 BMBC [29] 4 8 4 29.967.00 0.8750.126 26.327.78 0.7810.193 E2VID [32] 8 4 8 6.702.19 0.315 0.124 6.70 2.20 0.315 0.124 EDI [26] 4 4 8 18.76.53 0.574 0.244 18.8 6.88 0.579 0.274 Time Lens-syn (our) 4 4 4 30.575.01 0.903 0.067 28.98 5.09 0.873 0.086 Time Lens-real (ours) 4 4 4 32.494.60 0.927 0.048 30.57 5.08 0.900 0.069Figure 7: Qualitative results for the proposed method and its closes competitor DAIN [3] on our Dual Event and Color Camera Dataset test sequences: “Fountain Schaffhauserplatz” (top-left), “Fountain Bellevue” (bottom-left) “Water bomb” (top-right) and “Umbrella” (bottom-right). For each sequence, the figure shows interpolation results on the left (the animation can be viewed in Acrobat Reader) and close-up interpolation results on the right. The close-ups, show input left and right frame and intermediate interpolated frames. Table 4: Benchmarking on the test set of the High Speed Event and RGB camera (HS-ERGB) dataset. We report PSNR and SSIM for all sequences by skipping 5 and 7 frames respectively, and reconstructing the missing frames with each method. By design LEDVDI [15] can interpolate only 5 frames. Uses frames andUses events indicate if a method uses frames or events respectively. Color indicates whether a method works with color frames. For SSIM and PSNR the scores are averaged over the sequences. Best results are shown in bold and the second best are underlined. Method Uses frames Uses events Color PSNR SSIM PSNR SSIM Far-away sequences 5 frame skip 7 frames skips DAIN [3] 4 8 4 27.921.55 0.7800.141 27.131.75 0.7480.151 SuperSloMo [10] 4 8 4 25.666.24 0.727 0.221 24.16 5.20 0.692 0.199 RRIN [13] 4 8 4 25.265.81 0.738 0.196 23.73 4.74 0.703 0.170 BMBC [29] 4 8 4 25.626.13 0.742 0.202 24.13 4.99 0.710 0.175 LEDVDI [15] 4 4 8 12.501.74 0.393 0.174 n/a n/a Time Lens (ours) 4 4 4 33.132.10 0.877 0.092 32.31 2.27 0.869 0.110 Close planar sequences 5 frame skip 7 frames skips DAIN [3] 4 8 4 29.034.47 0.807 0.093 28.50 4.54 0:8010:096 SuperSloMo [10] 4 8 4 28.354.26 0.788 0.098 27.27 4.26 0:7750:099 RRIN [13] 4 8 4 28.694.17 0.813 0.083 27.46 4.24 0.800 0.084 BMBC [29] 4 8 4 29.224.45 0.8200.085 27.994.55 0.808 0.084 LEDVDI [15] 4 4 8 19.464.09 0.602 0.164 n/a n/a Time Lens (ours) 4 4 4 32.194.19 0.839 0.090 31.68 4.18 0.835 0.091 an up to 5.21 dB improvement over state-of-the-art frame-based and event-plus-frames-based methods on both synthetic and real datasets. In addition, we release the first High Speed Event and RGB (HS-ERGB) dataset, which aims at pushing the limits of existing interpola- tion approaches by establishing a new benchmark for both event- and frame-based video frame interpolation methods.5. Acknowledgement This work was supported by Huawei Zurich Research Center; by the National Centre of Competence in Re- search (NCCR) Robotics through the Swiss National Sci- ence Foundation (SNSF); the European Research Coun- cil (ERC) under the European Union’s Horizon 2020 re- search and innovation programme (Grant agreement No. 864042).6. Video Demonstration This PDF is accompanied with a video showing advan- tages of the proposed method compared to state-of-the-art frame-based methods published over recent months, as well as potential practical applications of the method. 7. Backbone network architecture Figure 8: Backbone hourglass network that we use in all modules of the proposed method. For all modules in the proposed method, we use the same backbone architecture which is an hourglass network with shortcut connections between the contracting and the ex- panding parts similar to [10] which we show in Fig. 8. 8. Additional Ablation Experiments 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 Percentage of locationsSynthesisWarped successiveWarp preceeding Figure 9: Percentage of pixels each interpolation method contributes on average to the final interpolation result for Vimeo90k (denoising) validation set. Note, that all meth- ods contribute almost equally to the final result and thus are equally important. Table 5: Importance of inter-frame events on Middlebury test set. To compute SSIM and PSNR, we skip one frame of the original video, reconstruct it and compare to the skipped frame. One version of the proposed method has access to the events synthesized from the skipped frame and another version does not have inter-frame information. We also show performance of frame-based SuperSloMo method [10], that is used in event simulator for reference. We highlight the best performing method. Method PSNR SSIM With inter-frame events (ours) 33.27 3.11 0.929 0.027 Without inter-frame events 29.03 4.85 0.866 0.111 SuperSloMo [10] 29.75 5.35 0.880 0.112 Importance of inter-frame events . To study the im- portance of additional information provided by events, we skip every second frame of the original video and attempt to reconstruct it using two versions of the proposed method.One version has access to the events synthesized from the skipped frame and another version does not have inter- frame information. As we can see from the Tab. 5, the former significantly outperforms the later by a margin of 4.24dB. Indeed this large improvements can be explained by the fact that the method with inter-frame events has im- plicit access to the ground truth image it tries to recon- struct, albeit in the form of asynchronous events. This high- lights that our network is able to efficiently decode the asyn- chronous intermediate events to recover the missing frame. Moreover, this shows that the addition of events has a sig- nificant impact on the final task performance, proving the usefulness of an event camera as an auxiliary sensor. Importance of each interpolation method. To study relative importance of synthesis-based andwarping-based interpolation methods, we compute the percentage of pixels that each method contribute on average to the final interpo- lation result for the Vimeo90k (denoising) validation dataset and show the result in Fig. 9. As it is clear from the figure, all the methods contribute almost equally to the final result and thus are all equally important. 1 2 3 4 5 6 7 Frame Index2224262830PSNR [dB] BMBC DAINRRIN SuperSlowMoOurs Figure 10: “Rope plot” showing interpolation quality as a function of distance from input boundary frames on High Quality Frames dataset. We skip all but every 7th frame and restore them using events and remaining frames. For each skip position, we compute average PSNR of the restored frame over entire dataset. We do not fine-tune the proposed and competing methods on the HQF dataset and simply use pre-trained models provided by the authors. Note, that the proposed method have the highest PSNR. Also, its PSNR decreases much slower than PSNR of other methods we move away from the input boundary frames. “Rope” plot. To study how the interpolation quality de- creases with the distance to the input frames, we skip all but every 7th frame in the input videos from the High Quality Frames dataset, restore them using our method and compare to the original frames. For each skipped frame position, we compute average PSNR of the restored frame over entire dataset and show results in Fig. 10. As clear from the fig-ure, the proposed method has the highest PSNR. Also, its PSNR decreases much slower than PSNR of the competing methods as we move away from the boundary frames. 9. Additional Benchmarking Results To makes sure that the fine-tuning does not af- fect our general conclusions, we fine-tuned the pro- posed method and RRIN method [13] on subset of High Quality Frames dataset and test them on the remaining part (“poster pillar 1”, “slow andfastdesk”, “bike bayhdr” and “desk” sequences). We choose RRIN method for this experiment, because it showed good perfor- mance across synthetic and real datasets and it is fairly sim- ple. As clear from the Tab. 6, after the fine-tuning, perfor- mance of the proposed method remained very strong com- pared to the RRIN method. 10. High Speed Events and RGB Dataset In this section we describe the sequences in the High- Speed Event and RGB (HS-ERGB) dataset. The commer- cially available DA VIS 346 [4] already allows the simul- taneous recording of events and grayscale frames, which are temporally and spatially synchronized. However, it has some shortcomings as the relatively low resolution of only 346260 pixels of both frames and events. This is far below the resolution of typical frame based consumer cam- eras. Additionally, the DA VIS 346 has a very limited dy- namic range of 55 db and a maximum frame of 40 FPS. Those properties render it not ideal for many event based methods, which aim to outperform traditional frame based cameras in certain applications. The setup described in [40] shows improvements in the resolution of frames and dy- namic range, but has a reduced event resolution instead. The lack of publicly available high resolution event and color frame datasets and of the shelf hardware motivated the de- velopment of our dual camera setup. It features high reso- lution, high frame rate, high dynamic range color frames combined with high resolution events. A comparison of our setup with the DA VIS346[4] and the setup with beam splitter in [40] is shown in 7. With this new setup we col- lect new High Speed Events and RGB (HS-ERGB) Dataset that we summarize in Tab. 8. We show several fragments from the dataset in Fig. 12. In the following paragraphs we describe temporal synchronization and spatial alignment of frame and event data that we performed for our dataset. Synchronization In our setup, two cameras are hard- ware synchronized through the use of external triggers. Each time the standard camera starts and ends exposure, a trigger is sent to the event camera which records an exter- nal trigger event with precise timestamp information. This information allows us to assign accurate timestamps to the standard frames, as well as group events during exposure orbetween consecutive frames. Alignment In our setup event and RGB cameras are ar- ranged in stereo configuration, therefore event and frame data in addition to temporal, require spatial alignment. We perform the alignment in three steps: (i)stereo calibration, (ii)rectification and (iii) feature-based global alignment. We first calibrate the cameras using a standard checker- board pattern. The recorded asynchronous events are con- verted to temporally aligned video reconstructions using E2VID[32, 33]. Finally, we find the intrinsic and extrin- sics by applying the stereo calibration tool Kalibr[24] to the video reconstructions and the standard frames recorded by the color camera. We then use the found intrinsics and ex- trinsics to rectify the events and frames. Due to the small baseline and similar fields of view (FoV), stereo rectification is usually sufficient to align the output of both sensors for scenes with a large average depth (>40 m ). This is illustrated in Fig. 11 (a). For close scenes, however, events and frames are mis- aligned (Fig. 11 (b)). For this reason we perform the sec- ond step of global alignment using a homography which we estimate by matching SIFT features [17] extracted on the standard frames and video reconstructions. The homog- raphy estimation also utilizes RANSAC to eliminate false matches. When the cameras are static, and the objects of interest move within a plane, this yields accurate alignment between the two sensors (Fig.11 (c)). References [1] Enhancing and experiencing spacetime resolution with videos and stills. In ICCP , pages 1–9. IEEE, 2009. 2 [2] Simon Baker, Daniel Scharstein, JP Lewis, Stefan Roth, Michael J Black, and Richard Szeliski. A database and evalu- ation methodology for optical flow. IJCV , 92(1):1–31, 2011. 6, 7 [3] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In CVPR , pages 3703–3712, 2019. 1, 5, 6, 7, 8 [4] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A 240 180 130 db 3 s la- tency global shutter spatiotemporal vision sensor. JSSC , 49(10):2333–2341, 2014. 2, 5, 10, 11 [5] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at once: Temporally adaptive multi-frame interpolation with ad- vanced motion modeling. arXiv preprint arXiv:2007.11762 , 2020. 2 [6] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carri ´o, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In CVPR , June 2020. 5, 6, 7 [7] Daniel Gehrig, Antonio Loquercio, Konstantinos G. Derpa- nis, and Davide Scaramuzza. End-to-end learning of repre- sentations for asynchronous event-based data. In Int. Conf. Comput. Vis. (ICCV) , 2019. 3Table 6: Results on High Quality Frames [35] with fine-tuning. Due to the time limitations, we only fine-tuned the pro- posed method and RRIN [13] method, that performed well across synthetic and real datasets. For evaluation, we used “poster pillar 1”, “slow andfastdesk”, “bike bayhdr” and “desk” sequences of the set and other sequences we used for the fine-tuning. For SSIM and PSNR, we show mean and one standard deviation across frames of all sequences. Method1 skip 3 skips PSNR SSIM PSNR SSIM RRIN [13] 28.62 5.51 0.839 0.132 25.36 5.70 0.750 0.173 Time Lens (Ours) 33.42 3.18 0.934 0.041 32.27 3.44 0.917 0.054 Table 7: Comparison of our HS-ERGB dataset against publicly available High Quality Frames (HQF) dataset, acquired by DA VIS 346 [4] and Guided Event Filtering (GEF) dataset, acquired by setup with DA VIS240 and RGB camera mounted with beam splitter [40]. Note, that in contrast to the previous datasets, the proposed dataset has high resolution of event data, and high frame rate. Also, it is the first dataset acquired by dual system with event and frame sensors arranged in stereo configuration. Frames Events FPS Dynamic Range, [dB] Resolution Color Dynamic Range, dB Resolution Sync. Aligned DA VIS 346 [4] 40 55 346 260 8 120 346 260 4 4 GEF[40] 35 60 24802048 4 120 240 180 4 4 HS-ERGB (Ours) 226 71.45 14401080 4 120 7201280 4 4 [8] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolu- tion of optical flow estimation with deep networks. In CVPR , pages 2462–2470, 2017. 2, 4 [9] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. In NIPS , pages 2017–2025, 2015. 2, 3 [10] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In CVPR , pages 9000–9008, 2018. 2, 3, 4, 5, 6, 7, 8, 9 [11] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. InCVPR , pages 3320–3329, 2020. 2 [12] Diederik P. Kingma and Jimmy L. Ba. Adam: A method for stochastic optimization. Int. Conf. Learn. Representations (ICLR) , 2015. 5 [13] Haopeng Li, Yuan Yuan, and Qi Wang. Video frame interpo- lation via residue refinement. In ICASSP 2020 , pages 2613– 2617. IEEE, 2020. 5, 6, 7, 8, 10, 11 [14] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A 128128 120 dB 15 s latency asynchronous temporal con- trast vision sensor. IEEE J. Solid-State Circuits , 43(2):566– 576, 2008. 2 [15] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. ECCV , 2020. 5, 6, 8 [16] Ziwei Liu, Raymond A Yeh, Xiaoou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In ICCV , pages 4463–4471, 2017. 2 [17] David G. Lowe. Distinctive image features from scale- invariant keypoints. Int. J. Comput. Vis. , 60(2):91–110, Nov. 2004. 5, 10[18] Simone Meyer, Abdelaziz Djelouah, Brian McWilliams, Alexander Sorkine-Hornung, Markus Gross, and Christo- pher Schroers. Phasenet for video frame interpolation. In CVPR , 2018. 2 [19] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In CVPR , July 2017. 7 [20] Simon Niklaus and Feng Liu. Context-aware synthesis for video frame interpolation. In CVPR , pages 1701–1710, 2018. 2, 3 [21] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In CVPR , pages 5437–5446, 2020. 2, 3, 4 [22] Simon Niklaus, Long Mai, and Feng Liu. Video frame inter- polation via adaptive convolution. In CVPR , 2017. 2 [23] Simon Niklaus, Long Mai, and Feng Liu. Video frame inter- polation via adaptive separable convolution. In ICCV , 2017. 2 [24] L. Oth, P. Furgale, L. Kneip, and R. Siegwart. Rolling shutter camera calibration. In CVPR , 2013. 10 [25] Avinash Paliwal and Nima Khademi Kalantari. Deep slow motion video reconstruction with hybrid imaging system. PAMI , 2020. 2 [26] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR , pages 6820–6829, 2019. 2, 6, 7 [27] Jiahao Pang, Wenxiu Sun, JS Ren, Chengxi Yang, and Qiong Yan. Cascade Residual Learning: A Two-stage Convolu- tional Neural Network for Stereo Matching. In ICCV , pages 887–895, 2017. 4 [28] Federico Paredes-Vall ´es and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruc- tion for event cameras via photometric constancy. CoRR , 2020. 2(a) far away scenes (b) misaligned close scenes (c) after global alignment Figure 11: Alignment of standard frames with events. Aggregated events (blue positive, red negative) are overlain with the standard frame. For scenes with sufficient depth (more than 40 m ) stereo rectification of both outputs yields accurate per-pixel alignment (a). However, for close scenes (b) events and frames are misaligned. In the absence of camera motion and motion in a plane, the views can be aligned with a global homography (c). [29] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost vol- ume for video interpolation. ECCV , 2020. 1, 2, 5, 6, 7, 8 [30] Pytorch web site. http://http://pytorch.org/ Accessed: 08 March 2019. 5 [31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an open event camera simulator. In Conf. on Robotics Learning (CoRL) , 2018. 5 [32] Henri Rebecq, Ren ´e Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In CVPR , pages 3857–3866, 2019. 2, 5, 7, 10 [33] Henri Rebecq, Ren ´e Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. TPAMI , 2019. 2, 6, 10 [34] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast im- age reconstruction with an event camera. In WACV , pages 156–163, 2020. 2 [35] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cam- eras. In ECCV , 2020. 6, 11 [36] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In CVPR , pages 8934–8943, 2018. 2 [37] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. ECCV , 2020. 2 [38] Lin Wang, Yo-Sung Ho, Kuk-Jin Yoon, et al. Event- based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In CVPR , pages 10081–10090, 2019. 2 [39] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility tostructural similarity. IEEE transactions on image processing , 13(4):600–612, 2004. 5 [40] Zihao Wang, Peiqi Duan, Oliver Cossairt, Aggelos Kat- saggelos, Tiejun Huang, and Boxin Shi. Joint filtering of in- tensity images and neuromorphic events for high-resolution noise-robust imaging. In CVPR , 2020. 5, 10, 11 [41] Zihao W Wang, Weixin Jiang, Kuan He, Boxin Shi, Aggelos Katsaggelos, and Oliver Cossairt. Event-driven video frame synthesis. In ICCV Workshops , pages 0–0, 2019. 2 [42] Chao-Yuan Wu, Nayan Singhal, and Philipp Krahenbuhl. Video compression through image interpolation. In ECCV , pages 416–431, 2018. 2 [43] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming- Hsuan Yang. Quadratic video interpolation. In NeurIPS , pages 1647–1656, 2019. 2, 3 [44] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. IJCV , 127(8):1106–1125, 2019. 2, 5, 6, 7 [45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR , pages 586–595, 2018. 5 [46] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow us- ing motion compensation. In ECCV , pages 0–0, 2018. 3Table 8: Overview of all sequences of the High Speed Event-RGB (HS-ERGB) dataset. Sequence Name Subset Camera Settings Description Close planar sequences Water bomb air (Fig. 12a) Train163 FPS, 1080 µsexposure, 1065 frames accelerating object, water splash Lighting match 150 FPS, 2972 µsexposure, 666 frames illumination change, fire Fountain Schaffhauserplatz 1 150 FPS, 977µsexposure, 1038 frames illumination change, fire Water bomb ETH 2 (Fig. 12c) 163 FPS, 323µsexposure, 3494 frames accelerating object, water splash Waving arms 163 FPS, 3476 µsexposure, 762 frames non-linear motion Popping air balloon Test150 FPS, 2972 µsexposure, 335 frames non-linear motion, object disappearance Confetti (Fig. 12e 150 FPS, 2972 µsexposure, 832 frames non-linear motion, periodic motion Spinning plate 150 FPS, 2971 µsexposure, 1789 frames non-linear motion, periodic motion Spinning umbrella 163 FPS, 3479 µsexposure, 763 frames non-linear motion Water bomb floor 1 (Fig. 12d) 160 FPS, 628µsexposure, 686 frames accelerating object, water splash Fountain Schaffhauserplatz 2 150 FPS, 977µsexposure, 1205 frames non-linear motion, water Fountain Bellevue 2 (Fig. 12b) 160 FPS, 480µsexposure, 1329 frames non-linear motion, water, periodic movement Water bomb ETH 1 163 FPS, 323µsexposure, 3700 frames accelerating object, water splash Candle (Fig. 12f) 160 FPS, 478µsexposure, 804 frames illumination change, non-linear motion Far-away sequences Kornhausbruecke letten x 1 Train163 FPS, 266µsexposure, 831 frames fast camera rotation around z-axis Kornhausbruecke rot x 5 163 FPS, 266µsexposure, 834 frames fast camera rotation around x-axis Kornhausbruecke rot x 6 163 FPS, 266µsexposure, 834 frames fast camera rotation around x-axis Kornhausbruecke rot y 3 163 FPS, 266µsexposure, 833 frames fast camera rotation around y-axis Kornhausbruecke rot y 4 163 FPS, 266µsexposure, 833 frames fast camera rotation around y-axis Kornhausbruecke rot z 1 163 FPS, 266µsexposure, 857 frames fast camera rotation around z-axis Kornhausbruecke rot z 2 163 FPS, 266µsexposure, 833 frames fast camera rotation around z-axis Sihl 4 163 FPS, 426µsexposure, 833 frames fast camera rotation around z-axis Tree 3 163 FPS, 978µsexposure, 832 frames camera rotation around z-axis Lake 4 163 FPS, 334µsexposure, 833 frames camera rotation around z-axis Lake 5 163 FPS, 275µsexposure, 833 frames camera rotation around z-axis Lake 7 163 FPS, 274µsexposure, 833 frames camera rotation around z-axis Lake 8 163 FPS, 274µsexposure, 832 frames camera rotation around z-axis Lake 9 163 FPS, 274µsexposure, 832 frames camera rotation around z-axis Bridge lake 4 163 FPS, 236µsexposure, 836 frames camera rotation around z-axis Bridge lake 5 163 FPS, 236µsexposure, 834 frames camera rotation around z-axis Bridge lake 6 163 FPS, 235µsexposure, 832 frames camera rotation around z-axis Bridge lake 7 163 FPS, 235µsexposure, 832 frames camera rotation around z-axis Bridge lake 8 163 FPS, 235µsexposure, 834 frames camera rotation around z-axis Kornhausbruecke letten random 4 Test163 FPS, 266µsexposure, 834 frames random camera movement Sihl 03 163 FPS, 426µsexposure, 834 frames camera rotation around z-axis Lake 01 163 FPS, 335µsexposure, 784 frames camera rotation around z-axis Lake 03 163 FPS, 334µsexposure, 833 frames camera rotation around z-axis Bridge lake 1 163 FPS, 237µsexposure, 833 frames camera rotation around z-axis Bridge lake 3 163 FPS, 236µsexposure, 834 frames camera rotation around z-axis(a) Water bomb air (b) Fountain Bellevue (c) Water bomb ETH 2 (d) Water bomb floor 1 (e) Confetti (f) Candle Figure 12: Example sequences of the HS-ERGB dataset. This figure contains animation that can be viewed in Acrobat Reader.