/ScarceNet

Primary LanguagePythonMIT LicenseMIT

ScarceNet

About

This is the source code for our paper

Chen Li, Gim Hee Lee. ScarceNet: Animal Pose Estimation with Scarce Annotations. In CVPR 2023.

In this paper, we aim to achieve accurate animal pose estimation with only a small set of labeled images and unlabeled images.

We design a pseudo label based framework to learn from scarce animal pose data. We first apply the small-loss trick to select a set of reliable pseudo labels. Despite its effectiveness, pseudo label selection by the small-loss trick tends to discard numerous high-loss samples. This results in high wastage since those discarded samples can still provide extra information for better discrimination. In view of this, we propose a reusable sample re-labeling step to further identify reusable samples from the high-loss samples via an agreement check and re-generate the corresponding pseudo labels for supervision. Lastly, we design a student-teacher framework to enforce consistency between the outputs of the student and teacher network.

For more details, please refer to our paper.

Dependencies

  1. Python 3.7
  2. Pytorch 1.7

Please refer to requirements.txt for more details on dependencies.

Download datasets

  • Clone this repository:

    https://github.com/chaneyddtt/ScarceNet.git
    
  • Download the AP10K dataset. Put the data and annotations under root/data/animalpose/.

  • Download the HRNet pretrained on the imagenet, and put under root/models/.

Test

Download our models and put them under the root/output folder. Test for the semi-supervised setting by running the command below. (We provide models trained with 5, 10, 15, 20 and 25 labels per category)

CUDA_VISIBLE_DEVICES=0 python tools/test.py --cfg experiments/ap10k/hrnet/w32_256x192_adam_lr1e-3.yaml --animalpose OUTPUT_DIR test TEST.MODEL_FILE output/output_part5_updatev2/model_best.pth MODEL.NAME pose_hrnet_part GPUS [0,]

You can also test for the transfer learning setting by running (modify DATASET.SUPERCATEGORY to test on different animal categories):

CUDA_VISIBLE_DEVICES=0 python tools/test.py --cfg experiments/ap10k/hrnet/w32_256x192_adam_lr1e-3.yaml --animalpose OUTPUT_DIR test TEST.MODEL_FILE output/output_part_updatev2_transfer/model_best.pth MODEL.NAME pose_hrnet_part GPUS [0,] DATASET.DATASET ap10k_test_category DATASET.SELECT_DATA True DATASET.SUPERCATEGORY "'deer',"

Train

If you do not want to train from scratch, you can download the models trained with the few labeled animal data, as well as the corresponding pseudo labels generated by these models. Move the folder pseudo labels under root/data/ and the pretrained models under root/output/. Train our model by running the command below. You can change the number of labels by setting 'LABEL_PER_CLASS' to 5, 10, 15, 20, 25 respectively. The pretrained model and output directory need to be changed acoordingly.

CUDA_VISIBLE_DEVICES=0,1 python tools/train_mt_part.py --cfg experiments/ap10k/hrnet/w32_256x192_adam_lr1e-3.yaml --animalpose --augment --pretrained output/output_animal_hrnet5_part/model_best.pth OUTPUT_DIR output_part5_updatev2 DATASET.DATASET ap10k_mt_v3 MODEL.NAME pose_hrnet_part TRAIN.BATCH_SIZE_PER_GPU 16 LABEL_PER_CLASS 5

You can also train from scratch. Firstly, train the model with the few labels by running:

CUDA_VISIBLE_DEVICES=0,1 python tools/train.py --cfg experiments/ap10k/hrnet/w32_256x192_adam_lr1e-3.yaml --animalpose OUTPUT_DIR output_animal_hrnet5_part DATASET.DATASET ap10k_fewshot MODEL.NAME pose_hrnet_part LABEL_PER_CLASS 5

Change the number of labels by setting 'LABEL_PER_CLASS' to 5, 10, 15, 20, 25 respectively, and the output directory need to be changed accordingly. Note that the labeled data are randomly selected from each category.

Create folder root/data/pseudo_labels/5shots/ and generate pseudo labels by running:

CUDA_VISIBLE_DEVICES=0 python tools/train_mt_part.py --cfg experiments/ap10k/hrnet/w32_256x192_adam_lr1e-3.yaml --animalpose --generate_pseudol --pretrained output/output_animal_hrnet5_part/model_best.pth OUTPUT_DIR test MODEL.NAME pose_hrnet_part TRAIN.BATCH_SIZE_PER_GPU 32 GPUS [0,] LABEL_PER_CLASS 5

Train the whole pipeline by running:

CUDA_VISIBLE_DEVICES=0,1 python tools/train_mt_part.py --cfg experiments/ap10k/hrnet/w32_256x192_adam_lr1e-3.yaml --animalpose --augment --pretrained output/output_animal_hrnet5_part/model_best.pth OUTPUT_DIR output_part5_updatev2 DATASET.DATASET ap10k_mt_v3 MODEL.NAME pose_hrnet_part TRAIN.BATCH_SIZE_PER_GPU 16 LABEL_PER_CLASS 5

The training steps for the transfer setting are similar. Firstly train with the labels from the Bovidae by running:

CUDA_VISIBLE_DEVICES=0,1 python tools/train.py --cfg experiments/ap10k/hrnet/w32_256x192_adam_lr1e-3.yaml --animalpose OUTPUT_DIR output_animal_hrnet_part_bovidae DATASET.DATASET ap10k MODEL.NAME pose_hrnet_part DATASET.SELECT_DATA True

Then create folder root/data/pseudo_labels/0shots, and generate pseudo labels by running:

CUDA_VISIBLE_DEVICES=0 python tools/train_mt_part.py --cfg experiments/ap10k/hrnet/w32_256x192_adam_lr1e-3.yaml --animalpose --generate_pseudol --pretrained output/output_animal_hrnet_part_bovidae/model_best.pth OUTPUT_DIR test MODEL.NAME pose_hrnet_part TRAIN.BATCH_SIZE_PER_GPU 32 GPUS [0,] LABEL_PER_CLASS 0

Lastly, train the whole pipeline by running:

CUDA_VISIBLE_DEVICES=0,1 python tools/train_mt_part.py --cfg experiments/ap10k/hrnet/w32_256x192_adam_lr1e-3.yaml --animalpose --augment --pretrained output/output_animal_hrnet_part_bovidae/model_best.pth --few_shot_setting OUTPUT_DIR output_part_updatev2_transfer DATASET.DATASET ap10k_mt_v3 MODEL.NAME pose_hrnet_part TRAIN.BATCH_SIZE_PER_GPU 16 LABEL_PER_CLASS 0 DATASET.SELECT_DATA True

Acknowledgements

The code for network architecture, data preprocessing, and evaluation are adapted from HRNet and AP-10K.