Repository files navigation

This repository contains the code for our ACM MM 2021 paper (oral), TalkNet, an active speaker detection model to detect 'whether the face in the screen is speaking or not?'. [Paper] [Video_English] [Video_Chinese].

A newdemo page. Thanks the contribution frommvoodarla !

• Awesome ASD: Papers about active speaker detection in last years.

• TalkNet in AVA-Activespeaker dataset: The code to preprocess the AVA-ActiveSpeaker dataset, train TalkNet in AVA train set and evaluate it in AVA val/test set.

• TalkNet in TalkSet and Columbia ASD dataset: The code to generate TalkSet, an ASD dataset in the wild, based on VoxCeleb2 and LRS3, train TalkNet in TalkSet and evaluate it in Columnbia ASD dataset.

• An ASD Demo with pretrained TalkNet model: An end-to-end script to detect and mark the speaking face by the pretrained TalkNet model.

Start from building the environment

conda create -n TalkNet python=3.7.9 anacondaconda activate TalkNetpip install -r requirement.txt

Start from the existing environment

pip install -r requirement.txt

The following script can be used to download and prepare the AVA dataset for training.

python trainTalkNet.py --dataPathAVA AVADataPath --download

AVADataPath is the folder you want to save the AVA dataset and its preprocessing outputs, the details can be found inhere . Please read them carefully.

Then you can train TalkNet in AVA end-to-end by using:

python trainTalkNet.py --dataPathAVA AVADataPath

exps/exps1/score.txt: output score file,exps/exp1/model/model_00xx.model: trained model,exps/exps1/val_res.csv: prediction for val set.

Our pretrained model performsmAP: 92.3 in validation set, you can check it by using:

python trainTalkNet.py --dataPathAVA AVADataPath --evaluation

The pretrained model will automaticly be downloaded intoTalkNet_ASD/pretrain_AVA.model. It performsmAP: 90.8 in the testing set.

We find that it is challenge to apply the model we trained in AVA for the videos not in AVA (Reason ishere, Q3.1). So we build TalkSet, an active speaker detection dataset in the wild, based onVoxCeleb2 andLRS3.

We do not plan to upload this dataset since we just modify it, instead of building it. InTalkSet folder we provide these.txt files to describe which files we used to generate the TalkSet and their ASD labels. You can generate thisTalkSet if you are interested to train an ASD model in the wild.

Also, we have provided our pretrained TalkNet model in TalkSet. You can evaluate it in Columbia ASD dataset or other raw videos in the wild.

A pretrain model in TalkSet will be download intoTalkNet_ASD/pretrain_TalkSet.model when using the following script:

python demoTalkNet.py --evalCol --colSavePath colDataPath

Also, Columnbia ASD dataset and the labels will be downloaded intocolDataPath. Finally you can get the following F1 result.

(This result is different from that in our paper because we train the model again, while the avg. F1 is very similar)

We build an end-to-end script to detect and extract the active speaker from the raw video by our pretrain model in TalkSet.

You can put the raw video (.mp4 and.avi are both fine) into thedemo folder, such as001.mp4.

python demoTalkNet.py --videoName 001

A pretrain model in TalkSet will be downloaded intoTalkNet_ASD/pretrain_TalkSet.model. The structure of the output reults can be found inhere.

You can get the output videodemo/001/pyavi/video_out.avi, which has marked the active speaker by green box and non-active speaker by red box.

If you want to evaluate by using cpu only, you can modifydemoTalkNet.py andtalkNet.py file: modify allcuda intocpu. Then replace line 83 in talkNet.py intoloadedState = torch.load(path,map_location=torch.device('cpu'))

Please cite the following if our paper or code is helpful to your research.

@inproceedings{tao2021someone,  title={Is Someone Speaking? Exploring Long-term Temporal Features for Audio-visual Active Speaker Detection},  author={Tao, Ruijie and Pan, Zexu and Das, Rohan Kumar and Qian, Xinyuan and Shou, Mike Zheng and Li, Haizhou},  booktitle = {Proceedings of the 29th ACM International Conference on Multimedia},  pages = {3927–3935},  year={2021}}

I have summaried some potentialFAQs. You can also check theissues in Github for other questions that I have answered.

This is my first open-source work, please let me know if I can future improve in this repositories or there is anything wrong in our work. Thanks for your support!

We study many useful projects in our codeing process, which includes:

The structure of the project layout and the audio encoder is learnt from thisrepository.

Demo for visulization is modified from thisrepository.

AVA data download code is learnt from thisrepository.

The model for the visual frontend is learnt from thisrepository.

Thanks for these authors to open source their code!

If you are interested to work on this topic and have some ideas to implement, I am glad to collaborate and contribute with my experiences & knowlegde in this topic. Please contact me withruijie.tao@u.nus.edu.