Home

Awesome

TBD

This is the official PyTorch implementation of our paper:

Tackling Background Distraction in Video Object Segmentation, ECCV 2022
Suhwan Cho, Heansung Lee, Minhyeok Lee, Chaewon Park, Sungjun Jang, Minjung Kim, Sangyoun Lee
Link: [ECCV] [arXiv]

<img src="https://user-images.githubusercontent.com/54178929/208470682-c36bfd92-db65-47ce-8e7e-7da51937686f.png" width=800>

You can also find other related papers at awesome-video-object-segmentation.

Abstract

In semi-supervised VOS, one of the main challenges is the existence of background distractors that have a similar appearance to the target objects. As comparing visual properties is a fundamental technique, visual distractions can severely lower the reliability of a system. To suppress the negative influence of background distractions, we propose three novel strategies: 1) a spatio-temporally diversified template construction scheme to prepare various object properties for reliable and stable prediction; 2) a learnable distance-scoring function to consider the temporal consistency of a video; 3) swap-and-attach data augmentation to provide hard training samples showing severe occlusions.

Preparation

1. Download COCO for network pre-training.

2. Download DAVIS and YouTube-VOS for network main training and testing.

3. Download Custom Split for YouTube-VOS training samples for network validation.

4. Replace dataset paths in "run.py" file with your dataset paths.

Training

1. Move to "run.py" file.

2. Check the training settings.

3. Run TBD training!!

python run.py --train

Testing

1. Move to "run.py" file.

2. Select a pre-trained model.

3. Run TBD testing!!

python run.py --test

Attachments

pre-trained model (davis)
pre-trained model (ytvos)
pre-computed results

Note

Code and models are only available for non-commercial research purposes.
If you have any questions, please feel free to contact me :)

E-mail: suhwanx@gmail.com