Fix reconstruct_and_export

Image3DProcessor.py

@@ -66,48 +66,46 @@ class Image3DProcessor:
         return image
     @torch.no_grad()
     def reconstruct_and_export(self, image):
-        # Ensure the input image is a NumPy array after preprocessing
-        if isinstance(image, Image.Image):
-            image = np.array(image)
-        elif isinstance(image, bytes):
-            image = np.array(Image.open(BytesIO(image)))
-
+        """
+        Passes image through model and outputs the reconstruction.
+        """
+        image= np.array(image)
         image_tensor = to_tensor(image).to(self.device)
         view_to_world_source, rot_transform_quats = get_source_camera_v2w_rmo_and_quats()
         view_to_world_source = view_to_world_source.to(self.device)
         rot_transform_quats = rot_transform_quats.to(self.device)
 
-        reconstruction_unactivated = self.model(
+        reconstruction_unactivated = model(
             image_tensor.unsqueeze(0).unsqueeze(0),
             view_to_world_source,
             rot_transform_quats,
             None,
             activate_output=False
         )
-
+        
         reconstruction = {k: v[0].contiguous() for k, v in reconstruction_unactivated.items()}
-        reconstruction["scaling"] = self.model.scaling_activation(reconstruction["scaling"])
-        reconstruction["opacity"] = self.model.opacity_activation(reconstruction["opacity"])
+        reconstruction["scaling"] = model.scaling_activation(reconstruction["scaling"])
+        reconstruction["opacity"] = model.opacity_activation(reconstruction["opacity"])
 
         # Render images in a loop
         world_view_transforms, full_proj_transforms, camera_centers = get_target_cameras()
         background = torch.tensor([1, 1, 1], dtype=torch.float32, device=self.device)
         loop_renders = []
-        t_to_512 = torchvision.transforms.Resize(512, interpolation=torchvision.transforms.InterpolationMode.LANCZOS)
+        t_to_512 = torchvision.transforms.Resize(512, interpolation=torchvision.transforms.InterpolationMode.NEAREST)
 
         for r_idx in range(world_view_transforms.shape[0]):
             rendered_image = render_predicted(
                 reconstruction,
-                world_view_transforms[r_idx].to(self.device),
-                full_proj_transforms[r_idx].to(self.device),
-                camera_centers[r_idx].to(self.device),
+                world_view_transforms[r_idx].to(device),
+                full_proj_transforms[r_idx].to(device),
+                camera_centers[r_idx].to(device),
                 background,
-                self.model_cfg,
+                model_cfg,
                 focals_pixels=None
             )["render"]
             rendered_image = t_to_512(rendered_image)
             loop_renders.append(torch.clamp(rendered_image * 255, 0.0, 255.0).detach().permute(1, 2, 0).cpu().numpy().astype(np.uint8))
-
+    
         # Save video to a file and load its content
         video_path = "loop_.mp4"
         imageio.mimsave(video_path, loop_renders, fps=25)