Demo code for Dynamic Label Graph Matching for Unsupervised Video Re-Identification in ICCV 2017.
We revised the evaluation protocol for the IDE on MARS dataset. In previous version, due to file traverse problem, which leads a different evaluation protocol, we achieve an extremely high performance (Unsupervised rank-1 65.2%, and supervised 75.8%) compared with other baselines in our cv-foundation version. We re-evaluate our perfomance under standard settings, the rank-1 is 36.8% for our unsupervised method, and the supervised upper bound is 56.2%. Please refer to the version on our website and github for latest results. PDF
-
a. You need to download our extracted features LOMO on BaiduYun and GoogleDrive or extract features by yourself. Put it under "data/" folder
-
b. You could run the demo_dgm.m and edit it to adjust for different datsets and different settings.
- LOMO on PRID-2011 and iLIDS-VID
Datasets | Rank@1 | Rank@5 | Rank@10 |
---|---|---|---|
#PRID-2011 | 73.1% | 92.5% | 96.7% |
#iLIDS-VID | 37.1% | 61.3% | 72.2% |
Notes: Due to the random graph generation, the results may be slighlty different.
-
a. You need to download our extracted features LOMO on BaiduYun and GoogleDrive or extract features by yourself. Put it under "data/" folder
-
b. You could run the demo_mars.m and edit it to adjust for different settings. Meanwhile, we could get the estimated labels.
-
c. With the estimated labels, we could re-arrange the dataset for IDE training.
-
d. Train IDE with our provided code follow the steps based on mxNet or try the baseline provided by Zhun Zhong.
We provide our trained models on BaiduYun and GoogleDrive for unsupervised and supervised baseline.
- On MARS dataset
Methods | Rank@1 | Rank@5 | mAP |
---|---|---|---|
#LOMO | 24.6% | 42.6% | 11.8% |
#IDE | 36.8% | 54.0% | 21.3% |
Please cite this paper in your publications if it helps your research:
@inproceedings{iccv17dgm,
title={Dynamic Label Graph Matching for Unsupervised Video Re-Identification},
author={Ye, Mang and Ma, Andy J and Zheng, Liang and Li, Jiawei and Yuen, Pong C.},
booktitle={ICCV},
year={2017},
}
Contact: [email protected]