Skip to content

Latest commit

 

History

History
148 lines (112 loc) · 8.41 KB

README.md

File metadata and controls

148 lines (112 loc) · 8.41 KB
Raw

iWildcam 2021 - UFAM Team

1st Place Solution to iWildcam 2021: Count the number of animals of each species present in a sequence of images

Requirements

Prepare an environment with python=3.8, tensorflow=2.3.1

Dependencies can be installed using the following command:

pip install -r requirements.txt

Data

Please refer to the iWildCam 2021 Github page for additional dataset details and download links.

Training

To train a classifier, use the script classification/train.py. Since we use multiple training stages, the script classification/multi_stage_train.py can be used to automate that:

python multi_stage_train.py --annotations_json=PATH_TO_BE_CONFIGURED/iwildcam2021_train_annotations.json \
    --megadetector_results_json=PATH_TO_BE_CONFIGURED/iwildcam2021train_originalimage_megadetector_v4.1_results_parsed.json \
    --dataset_dir=PATH_TO_BE_CONFIGURED/ \
    --model_name=efficientnet-b2 \
    --input_size=260 \
    --input_size_stage3=380 \
    --input_scale_mode=uint8 \
    --batch_size=32 \
    --lr_stage1=0.01 \
    --lr_stage2=0.01 \
    --lr_stage3=0.001 \
    --momentum=0.9 \
    --epochs_stage1=4 \
    --epochs_stage2=20 \
    --epochs_stage3=2 \
    --unfreeze_layers=18 \
    --label_smoothing=0.1 \
    --randaug_num_layers=6 \
    --randaug_magnitude=2 \
    --random_seed=42 \
    --model_dir=PATH_TO_BE_CONFIGURED/

The parameters can also be passed using a config file:

python multi_stage_train.py --flagfile=configs/efficientnet_b2_multicrop_bags_final_submission_training.config

Training using full image

By default, the training script trains models using bounding boxes crops from MegaDetector predictions. To use the full image, please specify --use_full_image.

Training using Balanced Group Softmax

To train a model using Balanced Group Softmax:

  1. Train a model, as usual, using classification/train.py or classification/multi_stage_train.py. By default, they will use the conventional softmax.
  2. Extract the base model weights using classification/save_base_model.py.
  3. Train a model by specifying --use_bags and also passsing the pretrained weights using --base_model_weights option. See classification/configs/efficientnet_b2_multicrop_bags_final_submission_training.config.

Prediction

Counting heuristic

For our final solution, we followed the competition benchmark counting heuristic: the maximum number of bounding boxes across any image in the sequence. We only counted bounding boxes from MegaDetectorV4 with confidence > 0.8.

This counting strategy limits us to predict only one species per sequence. We see our solution as a strong baseline for counting animals on camera trap sequences, but we do believe that the best solution should be based on tracking animals across images (Multi-object tracking), classifying each track, and counting them. We tried DeepSORT to track animals, but it was not as good as this heuristics for counting (see the next section).

Classification

The prediction assigned to each image is a weighted average of the predictions assigned to the bounding box with the highest score and to the full image (0.15 full image + 0.15 full image (flip) + 0.35 bbox + 0.35 bbox (flip)) as used by the winning solution of iWildCam 2020. To classify a sequence, we average the predictions of all nonempty images for each sequence.

Final submission

Both bbox and full image models of our final submission are based on EfficientNet-B2 using Balanced Group Softmax to handle the class imbalance problem in the dataset (see classification/configs folder). The bbox model was trained using all bounding boxes with confidence > 0.6. We run MegaDetector V4 to generate bounding boxes. The trained models used to generate predictions can be found here and the MegaDetector V4 bboxes for iWildCam 2021 are available here.

Use the script classification/predict_bbox_n_full.py to generate a submission:

python predict_bbox_n_full.py --annotations_json=PATH_TO_BE_CONFIGURED/iwildcam2021_train_annotations.json \
    --dataset_dir=PATH_TO_BE_CONFIGURED/ \
    --megadetector_results_json=PATH_TO_BE_CONFIGURED/iwildcam2021test_originalimage_megadetector_v4.1_results_parsed.json \
    --test_info_json=PATH_TO_BE_CONFIGURED/iwildcam2021_test_information.json \
    --submission_file_path=PATH_TO_BE_CONFIGURED/final_submission.csv \
    --model_name=efficientnet-b2 \
    --use_bags \
    --batch_size=16 \
    --input_size=380 \
    --input_scale_mode=uint8 \
    --ckpt_dir_full=PATH_TO_BE_CONFIGURED/fixefficientnet_b2_380x380_iwildcam_fulltrain_mdv4_fullimage_16mai_bags_mltstg/ \
    --ckpt_dir_bbox=PATH_TO_BE_CONFIGURED/fixefficientnet_b2_380x380_iwildcam_fulltrain_mdv4_multicrop_26mai_bags_mltstg/ \
    --use_flip_image \
    --ensemble_method=averaging \
    --megadetector_threshold=0.8

Other things that we tried

GPS coordinates

We tried to use the GPS coordinates and time of year (image timestamps) to train the Geo Prior model, but the model overfitted the iWildCam 2021 training set. We believe GPS coordinates can be useful for the problem (it worked very well for our iNat 2021 solution, for instance), but it's necessary to develop a model to deal with camera trap specificities, such as the fixed position.

This repository includes code to predict using a trained geo prior model that can be trained using our Tensorflow implementation. See the script classification/eval_main.py for predictions combined with geo priors. The Geo Prior model didn't learn using the original loss, so we had to replace it with the focal loss. But using only the classifier predictions was better than using them combined with geo priors.

DeepSORT to track animals

We tried to use DeepSORT to generate tracks over sequences and then classify each track by averaging the predictions of its bounding boxes list. We used the feature embedding from EfficientNet-B2 trained on bounding boxes. We also tried to use features from ReID models trained by the winner of ECCV TAO 2020 challenge, but they performed slightly worse than using EfficientNet-B2 features.

To extract features using EfficientNet-B2, use the script mot/generate_features.py:

python generate_features.py --model_name=efficientnet-b2 \
    --input_size=380 \
    --input_scale_mode=uint8 \
    --base_model_weights=PATH_TO_BE_CONFIGURED/efficientnet_b2_crop_25mai.h5 \
    --test_info_json=PATH_TO_BE_CONFIGURED/iwildcam2021_test_information.json \
    --dataset_dir=PATH_TO_BE_CONFIGURED \
    --megadetector_results_json=PATH_TO_BE_CONFIGURED/iwildcam2021test_originalimage_megadetector_v4.1_results_parsed.json \
    --min_confidence=0.9 \
    --random_seed=42 \
    --features_file=PATH_TO_BE_CONFIGURED/efficientnet_b2_crop_25mai_features.json

To track animals with DeepSORT use the script track_iwildcam.py from our DeepSORT repository. We kept DeepSORT code on a separate repository to avoid GPLv3 licensing conflicts.

Finally, to classify tracks and generate a submission, use the script classification/predict_track.py:

python predict_track.py --annotations_json=PATH_TO_BE_CONFIGURED/iwildcam2021_train_annotations.json \
    --dataset_dir=PATH_TO_BE_CONFIGURED \
    --megadetector_results_json=PATH_TO_BE_CONFIGURED/iwildcam2021test_originalimage_megadetector_v4.1_results_parsed.json \
    --test_info_json=PATH_TO_BE_CONFIGURED/iwildcam2021_test_information.json \
    --submission_file_path=PATH_TO_BE_CONFIGURED/deepsorttracks.csv \
    --model_name=efficientnet-b2 \
    --use_bags \
    --batch_size=32 \
    --input_size=380 \
    --input_scale_mode=uint8 \
    --ckpt_dir=PATH_TO_BE_CONFIGURED/fixefficientnet_b2_380x380_iwildcam_fulltrain_mdv4_multicrop_26mai_bags_mltstg/ \
    --tracks_file=PATH_TO_BE_CONFIGURED/efficientnet_b2_crop_25mai_tracks.json \
    --num_images_by_track=8

Contact

If you have any questions, feel free to contact Fagner Cunha (e-mail: fagner.cunha@icomp.ufam.edu.br) or Github issues.

License

Apache License 2.0