FathomNet - FGVC10 workshop at CVPR 2023

• Problem: The challenge is both to classify marine animals in a target image and assess if the image is from a different distribution relative to the training data.

Kaggle - Top 12/69 on private test

• Classification: ResNet101 or Dense as a backbone and 3 fully connected layers for classification head.
• Out of distribution: 2 approaches, first, use entropy to measure ood score. Second, enable dropout during inference time to sample many outputs and compute variance as ood score

Download Fathomnet2023 dataset

Train dataset

python3 download_images.py object_detection/train.json --outpath './datasets/train/'

Eval dataset

python3 download_images.py object_detection/eval.json --outpath './datasets/eval/'

Training

• Requirement: GPU

python3 train.py --arch resnet101 --dataset coco --save_dir ./saved_models/

Evaluation

• Models located: ./saved_models/coco/ Download models

python3 pred.py --arch resnet101