okankop / yowo

You Only Watch Once: A Unified CNN Architecture for Real-Time Spatiotemporal Action Localization

You Only Watch Once (YOWO)

PyTorch implementation of the article "You Only Watch Once: A Unified CNN Architecture for Real-Time Spatiotemporal Action Localization".

In this work, we present YOWO (You Only Watch Once), a unified CNN architecture for real-time spatiotemporal action localization in video stream. YOWO is a single-stage framework, the input is a clip consisting of several successive frames in a video, while the output predicts bounding box positions as well as corresponding class labels in current frame. Afterwards, with specific strategy, these detections can be linked together to generate Action Tubes in the whole video.

Since we do not separate human detection and action classification procedures, the whole network can be optimized by a joint loss in an end-to-end framework. We have carried out a series of comparative evaluations on two challenging representative datasets UCF101-24 and J-HMDB-21. Our approach outperforms the other state-of-the-art results while retaining real-time capability, providing 34 frames-per-second on 16-frames input clips and 62 frames-per-second on 8-frames input clips.

Installation

git clone https://github.com/wei-tim/YOWO.git
cd YOWO

Datasets

• UCF101-24: Download from here
• J-HMDB-21: Download from here

Modify the paths in ucf24.data and jhmdb21.data under cfg directory accordingly.

Download the dataset annotations from here.

Download backbone pretrained weights

• Darknet-19 weights can be downloaded via:

wget http://pjreddie.com/media/files/yolo.weights

• ResNeXt ve ResNet pretrained models can be downloaded from here.

NOTE: For JHMDB-21 trainings, HMDB-51 finetuned pretrained models should be used! (e.g. "resnext-101-kinetics-hmdb51_split1.pth").

• For resource efficient 3D CNN architectures (ShuffleNet, ShuffleNetv2, MobileNet, MobileNetv2), pretrained models can be downloaded from here.

Pretrained YOWO models

Pretrained models can be downloaded from here.

All materials (annotations and pretrained models) are also available in Baiduyun Disk: here with password 95mm

NOTE: With some extra tricks, YOWO achieves 87.5% and 76.7% frame_mAP for UCF101-24 and J-HMDB-21 datasets, respectively.

Running the code

• Example training bash scripts are provided in 'run_ucf101-24.sh' and 'run_jhmdb-21.sh'.
• UCF101-24 training:

python train.py --dataset ucf101-24 \
		--data_cfg cfg/ucf24.data \
		--cfg_file cfg/ucf24.cfg \
		--n_classes 24 \
		--backbone_3d resnext101 \
		--backbone_3d_weights weights/resnext-101-kinetics.pth \
		--backbone_2d darknet \
		--backbone_2d_weights weights/yolo.weights \

• J-HMDB-21 training:

python train.py --dataset jhmdb-21 \
		--data_cfg cfg/jhmdb21.data \
		--cfg_file cfg/jhmdb21.cfg \
		--n_classes 21 \
		--backbone_3d resnext101 \
		--backbone_3d_weights weights/resnext-101-kinetics-hmdb51_split1.pth \
		--freeze_backbone_3d \
		--backbone_2d darknet \
		--backbone_2d_weights weights/yolo.weights \
		--freeze_backbone_2d \

Validating the model

• After each validation, frame detections is recorded under 'jhmdb_detections' or 'ucf_detections'. From here, 'groundtruths_jhmdb.zip' and 'groundtruths_jhmdb.zip' should be downloaded and extracted to "evaluation/Object-Detection-Metrics". Then, run the following command to calculate frame_mAP.

python evaluation/Object-Detection-Metrics/pascalvoc.py --gtfolder PATH-TO-GROUNDTRUTHS-FOLDER --detfolder PATH-TO-DETECTIONS-FOLDER

• For video_mAP, 'run_video_mAP_ucf.sh' and 'run_video_mAP_jhmdb.sh' bash scripts can be used.

Citation

If you use this code or pre-trained models, please cite the following:

@InProceedings{kopuklu2019yowo,
title={You Only Watch Once: A Unified CNN Architecture for Real-Time Spatiotemporal Action Localization},
author={K{\"o}p{\"u}kl{\"u}, Okan and Wei, Xiangyu and Rigoll, Gerhard},
journal={arXiv preprint arXiv:1911.06644},
year={2019}
}

Acknowledgements

We thank Hang Xiao for releasing pytorch_yolo2 codebase, which we build our work on top.