3D ResNets for Action Recognition
Update (2020/4/13)
We published a paper on arXiv.
We uploaded the pretrained models described in this paper including ResNet-50 pretrained on the combined dataset with Kinetics-700 and Moments in Time.
Update (2020/4/10)
We significantly updated our scripts. If you want to use older versions to reproduce our CVPR2018 paper, you should use the scripts in the CVPR2018 branch.
This update includes as follows:
- Refactoring whole project
- Supporting the newer PyTorch versions
- Supporting distributed training
- Supporting training and testing on the Moments in Time dataset.
- Adding R(2+1)D models
- Uploading 3D ResNet models trained on the Kinetics-700, Moments in Time, and STAIR-Actions datasets
Summary
This is the PyTorch code for the following papers:
This code includes training, fine-tuning and testing on Kinetics, Moments in Time, ActivityNet, UCF-101, and HMDB-51.
Citation
If you use this code or pre-trained models, please cite the following:
@inproceedings{hara3dcnns,
author={Kensho Hara and Hirokatsu Kataoka and Yutaka Satoh},
title={Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet?},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
pages={6546--6555},
year={2018},
}
Pre-trained models
Pre-trained models are available here.
All models are trained on Kinetics-700 (K), Moments in Time (M), STAIR-Actions (S), or merged datasets of them (KM, KS, MS, KMS).
If you want to finetune the models on your dataset, you should specify the following options.
r3d18_K_200ep.pth: --model resnet --model_depth 18 --n_pretrain_classes 700
r3d18_KM_200ep.pth: --model resnet --model_depth 18 --n_pretrain_classes 1039
r3d34_K_200ep.pth: --model resnet --model_depth 34 --n_pretrain_classes 700
r3d34_KM_200ep.pth: --model resnet --model_depth 34 --n_pretrain_classes 1039
r3d50_K_200ep.pth: --model resnet --model_depth 50 --n_pretrain_classes 700
r3d50_KM_200ep.pth: --model resnet --model_depth 50 --n_pretrain_classes 1039
r3d50_KMS_200ep.pth: --model resnet --model_depth 50 --n_pretrain_classes 1139
r3d50_KS_200ep.pth: --model resnet --model_depth 50 --n_pretrain_classes 800
r3d50_M_200ep.pth: --model resnet --model_depth 50 --n_pretrain_classes 339
r3d50_MS_200ep.pth: --model resnet --model_depth 50 --n_pretrain_classes 439
r3d50_S_200ep.pth: --model resnet --model_depth 50 --n_pretrain_classes 100
r3d101_K_200ep.pth: --model resnet --model_depth 101 --n_pretrain_classes 700
r3d101_KM_200ep.pth: --model resnet --model_depth 101 --n_pretrain_classes 1039
r3d152_K_200ep.pth: --model resnet --model_depth 152 --n_pretrain_classes 700
r3d152_KM_200ep.pth: --model resnet --model_depth 152 --n_pretrain_classes 1039
r3d200_K_200ep.pth: --model resnet --model_depth 200 --n_pretrain_classes 700
r3d200_KM_200ep.pth: --model resnet --model_depth 200 --n_pretrain_classes 1039
Old pretrained models are still available here.
However, some modifications are required to use the old pretrained models in the current scripts.
Requirements
- PyTorch (ver. 0.4+ required)
conda install pytorch torchvision cudatoolkit=10.1 -c soumith
-
FFmpeg, FFprobe
-
Python 3
Preparation
ActivityNet
- Download videos using the official crawler.
- Convert from avi to jpg files using
util_scripts/generate_video_jpgs.py
python -m util_scripts.generate_video_jpgs mp4_video_dir_path jpg_video_dir_path activitynet
- Add fps infomartion into the json file
util_scripts/add_fps_into_activitynet_json.py
python -m util_scripts.add_fps_into_activitynet_json mp4_video_dir_path json_file_path
Kinetics
- Download videos using the official crawler.
- Locate test set in
video_directory/test.
- Locate test set in
- Convert from avi to jpg files using
util_scripts/generate_video_jpgs.py
python -m util_scripts.generate_video_jpgs mp4_video_dir_path jpg_video_dir_path kinetics
- Generate annotation file in json format similar to ActivityNet using
util_scripts/kinetics_json.py- The CSV files (kinetics_{train, val, test}.csv) are included in the crawler.
python -m util_scripts.kinetics_json csv_dir_path 700 jpg_video_dir_path jpg dst_json_path
UCF-101
- Download videos and train/test splits here.
- Convert from avi to jpg files using
util_scripts/generate_video_jpgs.py
python -m util_scripts.generate_video_jpgs avi_video_dir_path jpg_video_dir_path ucf101
- Generate annotation file in json format similar to ActivityNet using
util_scripts/ucf101_json.pyannotation_dir_pathincludes classInd.txt, trainlist0{1, 2, 3}.txt, testlist0{1, 2, 3}.txt
python -m util_scripts.ucf101_json annotation_dir_path jpg_video_dir_path dst_json_path
HMDB-51
- Download videos and train/test splits here.
- Convert from avi to jpg files using
util_scripts/generate_video_jpgs.py
python -m util_scripts.generate_video_jpgs avi_video_dir_path jpg_video_dir_path hmdb51
- Generate annotation file in json format similar to ActivityNet using
util_scripts/hmdb51_json.pyannotation_dir_pathincludes brush_hair_test_split1.txt, ...
python -m util_scripts.hmdb51_json annotation_dir_path jpg_video_dir_path dst_json_path
Running the code
Assume the structure of data directories is the following:
~/
data/
kinetics_videos/
jpg/
.../ (directories of class names)
.../ (directories of video names)
... (jpg files)
results/
save_100.pth
kinetics.json
Confirm all options.
python main.py -h
Train ResNets-50 on the Kinetics-700 dataset (700 classes) with 4 CPU threads (for data loading).
Batch size is 128.
Save models at every 5 epochs. All GPUs is used for the training. If you want a part of GPUs, use CUDA_VISIBLE_DEVICES=....
python main.py --root_path ~/data --video_path kinetics_videos/jpg --annotation_path kinetics.json \
--result_path results --dataset kinetics --model resnet \
--model_depth 50 --n_classes 700 --batch_size 128 --n_threads 4 --checkpoint 5
Continue Training from epoch 101. (~/data/results/save_100.pth is loaded.)
python main.py --root_path ~/data --video_path kinetics_videos/jpg --annotation_path kinetics.json \
--result_path results --dataset kinetics --resume_path results/save_100.pth \
--model_depth 50 --n_classes 700 --batch_size 128 --n_threads 4 --checkpoint 5
Calculate top-5 class probabilities of each video using a trained model (~/data/results/save_200.pth.)
Note that inference_batch_size should be small because actual batch size is calculated by inference_batch_size * (n_video_frames / inference_stride).
python main.py --root_path ~/data --video_path kinetics_videos/jpg --annotation_path kinetics.json \
--result_path results --dataset kinetics --resume_path results/save_200.pth \
--model_depth 50 --n_classes 700 --n_threads 4 --no_train --no_val --inference --output_topk 5 --inference_batch_size 1
Evaluate top-1 video accuracy of a recognition result (~/data/results/val.json).
python -m util_scripts.eval_accuracy ~/data/kinetics.json ~/data/results/val.json --subset val -k 1 --ignore
Fine-tune fc layers of a pretrained model (~/data/models/resnet-50-kinetics.pth) on UCF-101.
python main.py --root_path ~/data --video_path ucf101_videos/jpg --annotation_path ucf101_01.json \
--result_path results --dataset ucf101 --n_classes 101 --n_pretrain_classes 700 \
--pretrain_path models/resnet-50-kinetics.pth --ft_begin_module fc \
--model resnet --model_depth 50 --batch_size 128 --n_threads 4 --checkpoint 5
3 Dec 14, 2022
30 Sep 22, 2022
1 Jan 05, 2022
125 Dec 23, 2022
2 Feb 18, 2022
1.6k Jan 07, 2023
30 Dec 19, 2022
4.9k Dec 29, 2022
22 Feb 24, 2022
6 Nov 09, 2021
15 Dec 13, 2022
46 Nov 20, 2022
11 Dec 12, 2022
0 Sep 02, 2021
6 Nov 09, 2022
314 Nov 25, 2022
2.6k Jan 01, 2023
97 Nov 29, 2022
96 Jul 05, 2022
378 Jan 08, 2023