Discovering different environments in Animal Camera Traps
State-of-the-art algorithms show excellent performance ONLY when tested at the same Camera Trap location where they were trained but generalization to new locations is poor [1]. In this post, image clustering is utilized to discover/unveil different environments in camera trap data. Terra-Incognita 6GB sample camera trap data CCT20 Benchmark subset be downloaded from [2]*. Also, a 2GB Tarfile sample can be downloaded by running the download script from [3]. Camera Trap location 38 was selected and specifically images with dogs and cats labels. 200 random images were selected from cats pictures and another 200 from dogs. The clustering method is discussed in detail in a previous post [4]. Two clusters of almost equal size were generated from both samples.
Cats Cluster 1 and 2: 109,91
Dogs Cluster 1 and 2: 116,84
The main observation was the two clusters generated mainly two different environments day and night images. Also when looking at the farthest and closest to each cluster center, there is different zooming into the animal.
*Note: on downloading Metadata files for train/val/cis/trans splits (3MB), there will be a split schema called cis and trans locations. Cis locations is sample test and validation images from the same camera trap. While trans are images from a different one to measure model generalization.1-Cats image clustering results
1-Cats image clustering results
1.1-Cluster 1
1.1.1-Closest images to the cluster center
1.1.2-Farthest images from the cluster center
1.2-Cluster 2
1.2.1-Closest images to the cluster center
1.2.2-Farthest images from the cluster center
2-Dogs image clustering results
2.1-Cluster 1
2.1.1-Closest images to the cluster center
2.1.2-Farthest images from the cluster center
2.2-Cluster 2
2.2.1-Closest images to the cluster center
2.2.2-Farthest images from the cluster center
[1] https://arxiv.org/abs/1807.04975
[2] https://lila.science/datasets/caltech-camera-traps
[3]https://github.com/facebookresearch/DomainBed/blob/main/domainbed/scripts/download.py
