Awesome
Audio-Visual Question Answering (AVQA)
PyTorch code accompanies our CVPR 2022 paper:
Learning to Answer Questions in Dynamic Audio-Visual Scenarios (<font color="#FF0000">Oral Presentation</font>)
Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, Ji-Rong Wen and Di Hu
Resources: [Paper], [Supplementary], [Poster], [Video]
Project Homepage: https://gewu-lab.github.io/MUSIC-AVQA/
What's Audio-Visual Question Answering Task?
We focus on audio-visual question answering (AVQA) task, which aims to answer questions regarding different visual objects, sounds, and their associations in videos. The problem requires comprehensive multimodal understanding and spatio-temporal reasoning over audio-visual scenes.
<div align="center"> <img src="figs/avqa_white.png" width = "60%" /> </div>MUSIC-AVQA Dataset
The large-scale MUSIC-AVQA dataset of musical performance, which contains 45,867 question-answer pairs, distributed in 9,288 videos for over 150 hours. All QA pairs types are divided into 3 modal scenarios, which contain 9 question types and 33 question templates. Finally, as an open-ended problem of our AVQA tasks, all 42 kinds of answers constitute a set for selection.
-
QA examples
<div align="center"> <img src="figs/st_avqa_pairs.png" width = "100%" /> </div>
Model Overview
To solve the AVQA problem, we propose a spatio-temporal grounding model to achieve scene understanding and reasoning over audio and visual modalities. An overview of the proposed framework is illustrated in below figure.
<div align="center"> <img src="figs/framework.png" width = "90%" /> </div>Requirements
python3.6 +
pytorch1.6.0
tensorboardX
ffmpeg
numpy
Usage
-
Clone this repo
https://github.com/GeWu-Lab/MUSIC-AVQA_CVPR2022.git
-
Download data
Note: More information can be found in the DOWNLOAD Section of the project homepage
Raw videos
-
Baidu Drive (password: cvpr)
-
Real videos (36.67GB)
-
Synthetic videos (11.59GB)
-
-
Note
Please move all downloaded videos to a new folder, for example, create a new folder named MUSIC-AVQA-Videos, which contains 9,288 videos. eg., the downloaded videos will be in the
/data/video
folder.
Annotations (QA pairs, etc.)
-
Available for download at here
-
The annotation files are stored in JSON format. Each annotation file contains seven different keyword. And more detail see in Project Homepage
-
Data pre-processing
Extract audio waveforms from videos. The extracted audios will be in the
./data/audio
folder.moviepy library
is used to read videos and extract audios. (Note: If you are going to use the audio feature files provided by us, you can ignore this step.)cd ./feat_script python feat_script/extract_audio_cues/extract_audio.py
Extract frames from videos. The extracted frames will be in the
data/frames
folder.Pandas
andffmpeg
libraries are required.cd ./feat_script python extract_visual_frames/extract_frames_adaptive_script.py # or we can excute following statement to extract frames python extract_visual_frames/extract_frames.py
We also provide the extracted frames (1fps), which can be downloaded directly from <a href="https://pan.baidu.com/s/1c9gvJrf6oGXqHVtNiuOlZQ">Baidu Drive</a> (pwd: cvpr). But we thought it might be more convenient to execute the code above to extract video frames. Also, We should pay attention to whether the path in the code is correct.
-
Feature extraction
1). Audio feature.
TensorFlow1.4
andVGGish pretrained on AudioSet
is required.python feat_script/extract_audio_feat/audio_feature_extractor.py
Audio feature file also can be download from Google Drive or Baidu Drive (pwd: cvpr). VGGish feature shape: [T, 128].
2). 2D visual feature. Pretrained models library is required.
python feat_script/eatract_visual_feat/extract_rgb_feat.py
Features extracted using ResNet18 also can be downloaded from Google Drive or Baidu Drive (pwd: cvpr). ResNet18 feature shape: [T, 512]
3). 14x14 visual feature.
python feat_script/extract_visual_feat_14x14/extract_14x14_feat.py
14x14 features, too large to share ... but we can extract from raw video frames.
4). 3D visual feature.
The experiments in this paper do not use 3D snippet-level features, but we have carried out some additional experiments to show that 3D snippet-level features can effectively improve the performance of the model. If you want to use 3D snippet-level features, you can generate the relevant feature file by executing the following code, which can also be downloaded from Google Drive or Baidu Drive (pwd: cvpr) (973.9M). R(2+1)D feature shape: [T, 512].
python feat_script/eatract_visual_feat/extract_3d_feat.py
All the above feature files should be in
./data/feats
folder. -
Baseline Model
Training
python net_grd_baseline/main_qa_grd_baseline.py --mode train
Testing
python net_grd_baseline/main_qa_grd_baseline.py --mode test
-
Our Audio-Visual Spatial-Temporal Model
We provide trained models (path: ./net_grd_avst/avst_models/) and you can quickly test the results. Test results may vary slightly on different machines.
python net_grd_avst/main_avst.py --mode train \ --audio_dir = "path to your audio features" --video_res14x14_dir = "path to your visual res14x14 features"
Audio-Visual grounding generation
python grounding_gen/main_grd_gen.py
Training
python net_grd_avst/main_avst.py --mode train \ --audio_dir = "path to your audio features" --video_res14x14_dir = "path to your visual res14x14 features"
Testing
python net_grd_avst/main_avst.py --mode test \ --audio_dir = "path to your audio features" --video_res14x14_dir = "path to your visual res14x14 features"
Results
-
Audio-visual video question answering results of different methods on the test set of MUSIC-AVQA. The top-2 results are highlighted. Please see the citations in the [Paper] for comparison methods.
<div align="center"> <img src="figs/exp1.png" width = "100%" /> </div> -
Visualized spatio-temporal grounding results
We provide several visualized spatial grounding results. The heatmap indicates the location of sounding source. Through the spatial grounding results, the sounding objects are visually captured, which can facilitate the spatial reasoning.
Firstly,
./grounding_gen/models_grd_vis/
should be created.
<div align="center"> <img src="figs/visualize.png" width = "100%" /> </div>python grounding_gen/main_grd_gen_vis.py
Notice
We found some minor issues in dataset annotations, which are now fixed and placed in the 'json_update' folder. The experimental results based on the updated data are as follows:
<div align="center"> <img src="figs/exp3.png" width = "100%" /> </div>Citation
If you find this work useful, please consider citing it.
<pre><code> @ARTICLE{Li2022Learning, title = {Learning to Answer Questions in Dynamic Audio-Visual Scenarios}, author = {Guangyao li, Yake Wei, Yapeng Tian, Chenliang Xu, Ji-Rong Wen, Di Hu}, journal = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)}, year = {2022}, } </code></pre>Acknowledgement
This research was supported by Public Computing Cloud, Renmin University of China.
License
This project is released under the GNU General Public License v3.0.