andyvhuynh commited on
Commit
e9c4bbc
1 Parent(s): 000326f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +45 -1
README.md CHANGED
@@ -59,4 +59,48 @@ dataset_info:
59
  sequence:
60
  sequence:
61
  sequence: int64
62
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
59
  sequence:
60
  sequence:
61
  sequence: int64
62
+ ---
63
+
64
+ Nature Multi-View (NMV) Dataset Datacard
65
+ Dataset Overview
66
+
67
+ Name: Nature Multi-View Dataset (NMV)
68
+ Total Observations: 3,921,026
69
+ Total Images: Over 3 million ground-level and aerial image pairs
70
+ Species Covered: Over 6,000 native and introduced plant species
71
+ Geographic Focus: California, USA
72
+ Data Description
73
+
74
+ Ground-Level Images:
75
+ Sourced from iNaturalist open data on AWS.
76
+ Filters applied:
77
+ Vascular plants
78
+ Within California state boundaries
79
+ Observations dated from January 1, 2011, to September 27, 2023
80
+ Geographic uncertainty < 120 meters
81
+ Research-grade or in need of ID (excluding casual observations)
82
+ Availability of corresponding remote sensing imagery
83
+ Overlap with bio-climatic variables
84
+ Aerial Images:
85
+ Sourced from the 2018 National Agriculture Imagery Program (NAIP).
86
+ RGB-Infrared images, 256x256 pixels, 60 cm-per-pixel resolution.
87
+ Centered on the latitude and longitude of the iNaturalist observation.
88
+ Dataset Splits
89
+
90
+ Training Set:
91
+ Full Training Set: 1,755,602 observations, 3,307,025 images
92
+ Labeled Training Sets:
93
+ 20%: 334,383 observations, 390,908 images
94
+ 5%: 93,708 observations, 97,727 images
95
+ 1%: 19,371 observations, 19,545 images
96
+ 0.25%: 4,878 observations, 4,886 images
97
+ Validation Set:
98
+ 150,555 observations, 279,114 images
99
+ Test Set:
100
+ 182,618 observations, 334,887 images
101
+ Characteristics and Challenges
102
+
103
+ Long-Tail Distribution: The dataset exhibits a long-tail distribution common in natural world settings, making it a realistic benchmark for machine learning applications.
104
+ Geographic Bias: Reflects the geographic bias of citizen science data, with more observations from densely populated and visited regions like urban areas and National Parks.
105
+ Many-to-One Pairing: Multiple ground-level images are paired to the same aerial image, which may require special handling in machine learning models.
106
+ Self-Supervised Learning: Designed to encourage the development of self-supervised machine learning methods, unrestricted by label balance.