This repo holds codes of the paper: Temporal Lift Pooling for Continuous Sign Language Recognition.(ECCV 2022) [paper]
This repo is based on VAC (ICCV 2021). Many thanks for their great work!
Our training and inference procedure is modified from VAC (ICCV 2021). If you are familiar with VAC, you can play with TLP easily!
This project is implemented in Pytorch (>1.8). Thus please install Pytorch first.
ctcdecode==0.4 [parlance/ctcdecode]，for beam search decode.
sclite [kaldi-asr/kaldi], install kaldi tool to get sclite for evaluation. After installation, create a soft link toward the sclite:
mkdir ./software
ln -s PATH_TO_KALDI/tools/sctk-2.4.10/bin/sclite ./software/scliteSeanNaren/warp-ctc for ctc supervision.
The implementation for TLP is given in ./modules/tconv.py. You can choose to use TLP or max pooling or average pooling in line 77-79.
We notice that the kernel sizes of predictor and updater in lift pooling are set as 3 practically (which we wrongly report as 5 in the camera-ready version)
You can flexibly use TLP for other temporal tasks, e.g. action recogniton or video captioning.
You can choose any one of following datasets to verify the effectiveness of TLP.
Download the RWTH-PHOENIX-Weather 2014 Dataset [download link]. Our experiments based on phoenix-2014.v3.tar.gz.
After finishing dataset download, extract it. It is suggested to make a soft link toward downloaded dataset.
ln -s PATH_TO_DATASET/phoenix2014-release ./dataset/phoenix2014
The original image sequence is 210x260, we resize it to 256x256 for augmentation. Run the following command to generate gloss dict and resize image sequence.
cd ./preprocess
python dataset_preprocess.py --process-image --multiprocessingDownload the RWTH-PHOENIX-Weather 2014 Dataset [download link]
After finishing dataset download, extract it. It is suggested to make a soft link toward downloaded dataset.
ln -s PATH_TO_DATASET/PHOENIX-2014-T-release-v3/PHOENIX-2014-T ./dataset/phoenix2014-T
The original image sequence is 210x260, we resize it to 256x256 for augmentation. Run the following command to generate gloss dict and resize image sequence.
cd ./preprocess
python dataset_preprocess-T.py --process-image --multiprocessingThe results of TLP on CSL dataset is placed in the supplementary material.
Request the CSL Dataset from this website [download link]
After finishing dataset download, extract it. It is suggested to make a soft link toward downloaded dataset.
ln -s PATH_TO_DATASET ./dataset/CSL
The original image sequence is 1280x720, we resize it to 256x256 for augmentation. Run the following command to generate gloss dict and resize image sequence.
cd ./preprocess
python dataset_preprocess-CSL.py --process-image --multiprocessing| Backbone | Dev WER | Test WER | Pretrained model |
|---|---|---|---|
| Baseline | 21.2% | 22.3% | ------------------------------------------------------------ |
| ResNet18 | 19.7% | 20.8% | [Baidu] (passwd: tsa2) [Google Drive] |
| Backbone | Dev WER | Test WER | Pretrained model |
|---|---|---|---|
| Baseline | 21.1% | 22.8% | ------------------------------------------------------------ |
| ResNet18 | 19.4% | 21.2% | [Baidu] (passwd: c6cq) [Google Drive] |
| Backbone | WER | Pretrained model |
|---|---|---|
| Baseline | 7.3% | ------------------------------------------------------------ |
| ResNet18 | 1.8% | [Baidu] (passwd: e1md) [Google Drive] |
To evaluate the pretrained model, (if you test on the CSL dataset, you should first change the kernel size of predictor and updater in lift pooling to 5 with padding 2 (Line 12 and Line 19 in ./modules/tconv.py)), run the command below：
python main.py --device your_device --load-weights path_to_weight.pt --phase test
The priorities of configuration files are: command line > config file > default values of argparse. To train the SLR model on phoenix14, run the command below:
python main.py --device your_device
Note that you can choose the target dataset from phoenix2014/phoenix2014-T/CSL in line 3 in ./config/baseline.yaml.
If you want to train your model on the CSL dataset, please remind to (1) change the training epochs to 20, with steps of [10, 15] to decrease the learning rate in configs/ baseline.yaml. (2) change the kernel size of predictor and updater in lift pooling to 5 with padding 2 (Line 12 and Line 19 in ./modules/tconv.py).
If you find this repo useful in your research works, please consider citing:
@inproceedings{hu2022temporal,
title={Temporal lift pooling for continuous sign language recognition},
author={Hu, Lianyu and Gao, Liqing and Liu, Zekang and Feng, Wei},
booktitle={Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XXXV},
pages={511--527},
year={2022},
organization={Springer}
}