etimes2008 / vs-reid

Video Object Segmentation with Re-identification

Video Object Segmentation with Re-identification(VS-ReID) Pre-release

This repository holds the codes and models for the paper

Video Object Segmentation with Re-identification, Xiaoxiao Li, Yuankai Qi, Zhe Wang, Kai Chen, Ziwei Liu, Jianping Shi, Ping Luo, Xiaoou Tang, and Chen Change Loy
CVPR 2017 Workshop DAVIS Challenge on Video Object Segmentation 2017 (Winning Entry), Honolulu, Hawaii.

[Arxiv Preprint]

Prerequisites

---

• Python3
• PyTorch (Release version 0.4.0)

Get the code

---

Use git to clone this repository

git clone https://github.com/lxx1991/VS-ReID.git

Get the data, trained model, and pre-computed results

---

VS-Reid experiments on DAVIS datasets. After download, you need to remove the color-map from annotations.

It also needs the optical flow as input. In our paper, We use FlowNet2.0 to extract the optical flow for the whole dataset. For each pair of adjacent frames, we extract bidirectional optical flow, named as i.flo and (i+1).rflo.

Download the test-dev online finetuned propagation model.

We also provide the pre-computed classification and reid results for our model.

• classification result
• person-reid search result
• object-reid search result

The classification result is the ImageNet classification score of each foreground object generated by ResNet-101

The person-reid search result is pre-computed by Person Search, and object-reid search result is generated by a Faster R-CNN detector and retrained "Person Search-Similar" network.

Documentation

---

The directory is structured as follows:

.
├── data
|   └── DAVIS
│       ├── Annotations
|       |   └── ... 
│       ├── JPEGImages
|       |   └── ... 
|       ├── Flow
│       │   └── 480p
|       |       ├── aerobatics
|       |       |   ├── 00000.flo
|       |       |   ├── 00001.rflo
|       |       |   ├── 00001.flo
|       |       |   └── ... 
|       |       └── ... 
│       ├── ObjectSearch
│       │   └── 480p
|       |       └──test-dev.pkl
│       ├── PersonSearch
│       │   └── 480p
|       |       └──test-dev.pkl
│       ├── Class
│       │   └── 480p
|       |       └──test-dev.pkl      
│       └── ...
├── models                  
│   └── MP2S.pth.tar
└── ...

Usage

---

Single gpu

python3 davis_test.py test-dev configs/test_config.py --output OUT_DIR_NAME --cache CACHE_DIR_NAME --gpu 0

Multiple gpus

./run.sh OUT_DIR_NAME CACHE_DIR_NAME GPU_NUM

Citation

Please cite the following paper if you feel this repository useful.

@inproceedings{li2017video,
  author    = {Li, Xiaoxiao and Qi, Yuankai and Wang, Zhe and Chen, Kai and Liu, Ziwei and Shi, Jianping and Luo, Ping and Tang, Xiaoou and Loy, Chen Change},
  title     = {Video Object Segmentation with Re-identification},
  booktitle   = {The 2017 DAVIS Challenge on Video Object Segmentation - CVPR Workshops},
  year      = {2017},
}

TODOs

---

• Update README
• ReID evaluation script
• Training Code

Contact

---

For any question, feel free to contact

Xiaoxiao Li : [email protected]