This is the official implementation of the Wnet paper:
Audio-Guided video object segmentation is a challenging problem in visual analysis and editing, which automatically separates foreground objects from the background in a video sequence according to the referring audio expressions. However, the existing referring video object segmentation works mainly focus on the guidance of text-based referring expressions, due to the lack of modeling the semantic representations of audio-video interaction contents. In this paper, we consider the problem of audio-guided video semantic segmentation from the viewpoint of end-to-end denoising encoder-decoder network learning. The extensive experiments show the effectiveness of our method.
First, clone the repo locally:
git clone https://github.com/asudahkzj/Wnet.git
Then, install PyTorch 1.8 and torchvision 0.9:
conda install pytorch==1.8.0 torchvision==0.9.0
Install pytorch_wavelets
git clone https://github.com/fbcotter/pytorch_wavelets
cd pytorch_wavelets
pip install .
Install pycocotools
conda install cython scipy
pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
pip install git+https://github.com/youtubevos/cocoapi.git#"egg=pycocotools&subdirectory=PythonAPI"
If you encounter the problem of missing ytvos.py file, you can manually download the file from here and put it in the installed pycocotools folder.
Compile DCN module(requires GCC>=5.3, cuda>=10.0)
cd models/dcn
python setup.py build_ext --inplace
Download and extract 2021 version of Refer-Youtube-VOS train images from RVOS. Follow the instructions here to download A2D-Sentences and JHMDB-Sentences dataset. The new audio dataset (AVOS) is also open. You need to extract MFCC features from audio files and convert video files in A2D into image frames. For extracting MFCC features, you can refer to here.
Then, organize the files as follows:
Wnet/data
├── rvos/
│ ├── ann/
| | └── *.json (annotation files)
│ ├── train/
| │ ├── JPEGImages/
| │ └── Annotations/
│ └── meta_expressions/train/meta_expressions.json
├── a2d/
| ├── Release/
| │ ├── videoset.csv
| │ ├── clips320/
| │ └── pngs320/ (image frames extracted from videos in clips320/)
| ├── a2d_annotation_with_instances/
| └── a2d_annotation_info.txt
├── jhmdb/
| ├── Rename_Images/
| ├── puppet_mask/
| ├── jhmdb_annotation.txt
| └── video_info.json
├── rvos_audio_feature/
| └── *.npy (mfcc features extracted from rvos audio file)
└── a2d_j_audio_feature/
└── *.npy (mfcc features extracted from a2d/jhmdb audio file)
*The files a2d_annotation_info.txt and video_info.json can be downloaded here.
Download the pretrained DETR models OneDrive on COCO and save it to the pretrained path.
For AVOS dataset (which contains the videos and audios of RVOS, A2D-Sentences and JHMDB-Sentences, and JHMDB-Sentences dataset is only for evaluation):
python -m torch.distributed.launch --nproc_per_node=4 --use_env main.py --backbone resnet101/50 --dataset_file avos --ytvos_path /path/to/ytvos --masks --pretrained_weights /path/to/pretrained_path
For RVOS dataset:
python -m torch.distributed.launch --nproc_per_node=4 --use_env main.py --backbone resnet101/50 --dataset_file ytvos --ytvos_path /path/to/ytvos --masks --pretrained_weights /path/to/pretrained_path
For A2D-Sentences dataset:
python -m torch.distributed.launch --nproc_per_node=4 --use_env main.py --backbone resnet101/50 --dataset_file a2d --num_frames 8 --num_queries 8 --masks --pretrained_weights /path/to/pretrained_path
For AVOS dataset, we need to test three datasets separately:
python inference_rvos.py --masks --model_path /path/to/model_weights --save_path /path/to/results.json
python evaluate/evaluate.py /path/to/results.json
python inference_a2d.py --masks --model_path /path/to/model_weights
python inference_jh.py --masks --model_path /path/to/model_weights
For RVOS dataset:
python inference_rvos.py --masks --model_path /path/to/model_weights --save_path /path/to/results.json
python evaluate/evaluate.py /path/to/results.json
For A2D-Sentences dataset:
python inference_a2d.py --masks --model_path /path/to/model_weights --num_frames 8 --num_queries 8
For JHMDB-Sentences dataset (directly using the model trained on A2D-Sentences):
python inference_jh.py --masks --model_path /path/to/model_weights --num_frames 8 --num_queries 8
We provide Wnet models trained from the AVOS dataset, which contains the videos of RVOS, A2D-Sentences and JHMDB-Sentences.
| Name | Backbone | J | F | J&F | Chenkpoint |
|---|---|---|---|---|---|
| Wnet | ResNet-50 | 43.0 | 45.0 | 44.0 | Link |
Please consider citing our work in your publications if the project helps your research:
@InProceedings{Pan_2022_CVPR,
author = {Pan, Wenwen and Shi, Haonan and Zhao, Zhou and Zhu, Jieming and He, Xiuqiang and Pan, Zhigeng and Gao, Lianli and Yu, Jun and Wu, Fei and Tian, Qi},
title = {Wnet: Audio-Guided Video Object Segmentation via Wavelet-Based Cross-Modal Denoising Networks},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2022},
pages = {1320-1331}
}