Code for CVPR2021 ChaLearn Challenge of ISLR

Our Team Ranked 3rd in 2021 ChaLearn LAP ISLR CVPR Challenge.

1. Run Docker

Run the docker first and then execute all the following steps.

docker pull rhythmblue/chalearn2021:v1
docker run -it --rm --shm-size=16G -v /xxx/ChaLearn-2021-ISLR-Challenge:/ChaLearn-2021-ISLR-Challenge rhythmblue/chalearn2021:v1
cd /ChaLearn-2021-ISLR-Challenge

2. Preprocessed data

• Download the preprocessed test data at GoogleDrive and organize them as follows,

data/    
└── AUTSL
    ├── autsl_01_split.json
    ├── autsl_final.json 
    ├── flow          
    │   └── flow.zip  
    ├── jpg_face    
    │   └── face.zip  
    ├── jpg_left_hand   
    │   └── lhand.zip  
    ├── jpg_right_hand  
    │   └── rhand.zip 
    ├── jpg_video 
    │   └── full.zip
    └── Keypoints_2d_mmpose
        └── kps.zip    

• Unzip

cd /ChaLearn-2021-ISLR-Challenge
sh 1_unzip.sh

• Download the pretrained models and weights GoogleDrive, and organize them as follows,

code/                                 
├── pre_trained
├── weights
└── ...   

3. Inference

In this step, we need to run the commands in 2_inference.sh in order.

cd /ChaLearn-2021-ISLR-Challenge
sh 2_inference.sh

4. Ensemble

Run the script to generate the .csv file.

cd /ChaLearn-2021-ISLR-Challenge
sh 3_ensemble.sh

5. Final Submission

• RGB: predictions_fusion-final-RGB.csv
• RGBD: predictions_fusion-final-RGBD.csv

Training (If needed)

The detailed training script are provided in script/train.sh

Precessing the data by yourself (If needed)

• Unzip the original videos in the raw folder.

preprocessing/data/AUTSL
├── first               
│   ├── test            
│   ├── train           
│   └── val             
├── flow                
│   ├── test            
│   ├── train           
│   └── val             
├── image               
│   ├── test            
│   ├── train           
│   └── val             
├── label               
│   ├── test_random.csv 
│   ├── train_labels.csv
│   └── val_random.csv  
└── raw                 
    ├── test            
    ├── train           
    └── val             

• Run docker

docker pull rhythmblue/openpose:cuda11.1-cudnn8-v1
docker run -it docker run -it --rm --shm-size=16G -v /xxx/ChaLearn-2021-ISLR-Challenge:/ChaLearn-2021-ISLR-Challenge rhythmblue/openpose:cuda11.1-cudnn8
cd /ChaLearn-2021-ISLR-Challenge

• extract the first frame of each video to help localize signers.

cd preprocessing
python preparation/1.gen_list.py
python preparation/2.extract_first.py data/AUTSL/raw data/AUTSL/first

• localize signer and crop the video. Pre-trained weights preprocessing/preparation/weights/pose_hrnet_w48_384x288.pth and preprocessing/preparation/pre_data/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth can be downloaded at GoogleDrive

cd preprocessing
python preparation/3.localize_signer.py
python preparation/4.gen_ffmpeg_list.py
sh ffmpeg_folder_train.sh
sh ffmpeg_folder_val.sh
sh ffmpeg_folder_test.sh
sh ffmpeg_train.sh
sh ffmpeg_val.sh
sh ffmpeg_test.sh

• extract optical flow

cd preprocessing
python preparation/5.compute_flow.py

• extract keypoints with DarkPose-Wholebody in MMPose. Pre-trained weights of mmpose/models/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth and mmpose/models/hrnet_w48_coco_wholebody_384x288_dark-f5726563_20200918.pth can be downloaded at GoogleDrive

python Stage1_top_down_SL_pose_video_AUTSL.py \
    demo/mmdetection_cfg/faster_rcnn_r50_fpn_1x_coco.py models/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
    configs/top_down/darkpose/coco-wholebody/hrnet_w48_coco_wholebody_384x288_dark.py \
    models/hrnet_w48_coco_wholebody_384x288_dark-f5726563_20200918.pth 100

• Arrange data

python Stage2_All_Prepare_txt_file_w_val.py
python Stage3_n_frames_ucf101_hmdb51.py /data/user/AUTSL/jpg_video
python Stage4_AUTSL_json.py All_new
python Stage8_Crop_Hand.py
python Stage9_Crop_Face.py

Hao Zhou, zhouh156(AT)mail.ustc.edu.cn