AOT (Associating Objects with Transformers for Video Object Segmentation) in PyTorch

A modular reference PyTorch implementation of Associating Objects with Transformers for Video Object Segmentation (NeurIPS 2021, Score 8/8/7/8). [OpenReview][PDF]

Benchmark examples:

General examples (Messi and Kobe):

Highlights

High performance: up to 85.5% (R50-AOTL) on YouTube-VOS 2018 and 82.1% (SwinB-AOTL) on DAVIS-2017 Test-dev under standard settings (without any test-time augmentation and post processing).
High efficiency: up to 51fps (AOTT) on DAVIS-2017 (480p) even with 10 objects and 41fps on YouTube-VOS (1.3x480p). AOT can process multiple objects (less than a pre-defined number, 10 is the default) as efficiently as processing a single object. This project also supports inferring any number of objects together within a video by automatic separation and aggregation.
Multi-GPU training and inference
Mixed precision training and inference
Test-time augmentation: multi-scale and flipping augmentations are supported.

Requirements

Python3
pytorch >= 1.7.0 and torchvision
opencv-python
Pillow
Pytorch Correlation (Recommend to install from source instead of using pip. The project can also work without this moduel but will lose some efficiency of the short-term attention.)

Optional:

scikit-image (if you want to run our Demo, please install)

Model Zoo and Results

Pre-trained models, benckmark scores, and pre-computed results reproduced by this project can be found in MODEL_ZOO.md.

Demo - Panoptic Propagation

We provide a simple demo to demonstrate AOT's effectiveness. The demo will propagate more than 40 objects, including semantic regions (like sky) and instances (like person), together within a single complex scenario and predict its video panoptic segmentation.

To run the demo, download the checkpoint of R50-AOTL into pretrain_models, and then run:

python tools/demo.py

which will predict the given scenarios in the resolution of 1.3x480p. You can also run this demo with other AOTs (MODEL_ZOO.md) by setting --model (model type) and --ckpt_path (checkpoint path).

Two scenarios from VSPW are supplied in datasets/Demo:

1001_3iEIq5HBY1s: 44 objects. 1080P.
1007_YCTBBdbKSSg: 43 objects. 1080P.

Results:

Getting Started

Prepare a valid environment follow the requirements.
Prepare datasets:

Please follow the below instruction to prepare datasets in each corresponding folder.
- Static
  
  datasets/Static: pre-training dataset with static images. Guidance can be found in AFB-URR, which we referred to in the implementation of the pre-training.
- YouTube-VOS
  
  A commonly-used large-scale VOS dataset.
  
  datasets/YTB/2019: version 2019, download link. train is required for training. valid (6fps) and valid_all_frames (30fps, optional) are used for evaluation.
  
  datasets/YTB/2018: version 2018, download link. Only valid (6fps) and valid_all_frames (30fps, optional) are required for this project and used for evaluation.
- DAVIS
  
  A commonly-used small-scale VOS dataset.
  
  datasets/DAVIS: TrainVal (480p) contains both the training and validation split. Test-Dev (480p) contains the Test-dev split. The full-resolution version is also supported for training and evaluation but not required.
Prepare ImageNet pre-trained encoders

Select and download below checkpoints into pretrain_models:
- MobileNet-V2 (default encoder)
- MobileNet-V3
- ResNet-50
- ResNet-101
- ResNeSt-50
- ResNeSt-101
- Swin-Base
The current default training configs are not optimized for encoders larger than ResNet-50. If you want to use larger encoders, we recommend early stopping the main-training stage at 80,000 iterations (100,000 in default) to avoid over-fitting on the seen classes of YouTube-VOS.
Training and Evaluation

The example script will train AOTT with 2 stages using 4 GPUs and auto-mixed precision (--amp). The first stage is a pre-training stage using Static dataset, and the second stage is a main-training stage, which uses both YouTube-VOS 2019 train and DAVIS-2017 train for training, resulting in a model that can generalize to different domains (YouTube-VOS and DAVIS) and different frame rates (6fps, 24fps, and 30fps).

Notably, you can use only the YouTube-VOS 2019 train split in the second stage by changing pre_ytb_dav to pre_ytb, which leads to better YouTube-VOS performance on unseen classes. Besides, if you don't want to do the first stage, you can start the training from stage ytb, but the performance will drop about 1~2% absolutely.

After the training is finished (about 0.6 days for each stage with 4 Tesla V100 GPUs), the example script will evaluate the model on YouTube-VOS and DAVIS, and the results will be packed into Zip files. For calculating scores, please use official YouTube-VOS servers (2018 server and 2019 server), official DAVIS toolkit (for Val), and official DAVIS server (for Test-dev).

Adding your own dataset

Coming

Troubleshooting

Waiting

TODO

Citations

Please consider citing the related paper(s) in your publications if it helps your research.

@inproceedings{yang2021aot,
  title={Associating Objects with Transformers for Video Object Segmentation},
  author={Yang, Zongxin and Wei, Yunchao and Yang, Yi},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2021}
}

License

This project is released under the BSD-3-Clause license. See LICENSE for additional details.

mkg1204 / aot-benchmark