What's new: We updated the name of the dataset from 'Duke' to 'DukeMTMC-reID', added the original license from DukeMTMC and removed the redistribution limitation.
DukeMTMC-reID is a subset of the DukeMTMC for image-based re-identification, in the format of the Market-1501 dataset. The original dataset contains 85-minute high-resolution videos from 8 different cameras. Hand-drawn pedestrain bounding boxes are available.
We crop pedestrain images from the videos every 120 frames, yielding in total 36,411 bounding boxes with IDs. There are 1,404 identities appearing in more than two cameras and 408 identities (distractor ID) who appear in only one camera. We randomly select 702 IDs as the training set and the remaining 702 IDs as the testing set. In the testing set, we pick one query image for each ID in each camera and put the remaining images in the gallery.
As a result, we get 16,522 training images of 702 identities, 2,228 query images of the other 702 identities and 17,661 gallery images.
| File | Description |
|---|---|
| /bounding_box_test | The gallery images. We retrieve a query from this image pool. |
| /bounding_box_train | The training images. This dir contains the images from 702 different identities. |
| /query | The query images. Each of them is from different identities in different cameras. |
Naming Rule of the images In bbox "0005_c2_f0046985.jpg", "0005" is the identity. "c2" means the image from Camera 2. "f0046985" is the 46985th frame in the video of Camera 2.
Please follow the LICENSE_DukeMTMC-reID. You are free to share, create and adapt the DukeMTMC-reID dataset, in the manner specified in the license.
We also include the LICENSE_DukeMTMC. If you want to share, create and adapt the DukeMTMC dataset, please follow this license.
The DukeMTMC-reID evaluation code is under the MIT License.
You can download the DukeMTMC-reID dataset from GoogleDriver or (BaiduYun password: chu1).
Some unzip tools on Windows may meet some problems. Please check that you have the following files after unzip:
If download links are unavailable, please don't hesitate to contact me to update links. Thank you.
Figure. The image distribution of DukeMTMC-reID training set. We note that the median of images per ID is 20. But some ID may contain lots of images, which may comprise some algorithms. (For example, ID 5388 contains 426 images.)
Thank Xun for suggestions.
To evaluate, you need to calculate your gallery and query feature (i.e., 17661x2048 and 2228x2048 matrix) and save them in advance. Then download the codes in this repository. You just need to change the image path and the feature path in the evaluation_res_duke_fast.m and run it to evaluate.
| Methods | Rank@1 | mAP | Reference |
|---|---|---|---|
| BoW+kissme | 25.13% | 12.17% | "Scalable person re-identification: a benchmark", Zheng Liang, Shen Liyue, Tian Lu, Wang Shengjin, Wang Jingdong and Tian, Qi, ICCV 2015 [project] |
| LOMO+XQDA | 30.75% | 17.04% | "Person Re-identification by Local Maximal Occurrence Representation and Metric Learning", Liao Shengcai, Hu Yang, Zhu Xiangyu and Li Stan Z, CVPR 2015 [project] |
| Basel. | 65.22% | 44.99% | "Person Re-identification: Past, Present and Future", Zheng Liang, Yi Yang, and Alexander G. Hauptmann, arXiv:1610.02984 [code] |
| Basel. + LSRO | 67.68% | 47.13% | "Unlabeled Samples Generated by GAN Improve the Person Re-identification Baseline in vitro", Zheng Zhedong, Zheng Liang and Yang Yi, ICCV 2107 |
| Basel. + OIM | 68.1% | - | "Joint Detection and Identification Feature Learning for Person Search", Xiao Tong, Li Shuang , Wang Bochao , Lin Liang and Wang Xiaogang, CVPR 2017 |
| Verif + Identif | 68.9% | 49.3% | "A Discriminatively Learned Cnn Embedding for Person Re-identification", Zheng, Zhedong, Liang Zheng, and Yi Yang, arXiv:1611.05666. [code] |
| APR | 70.69% | 51.88% | "Improving person re-identification by attribute and identity learning", Lin Yutian, Zheng Liang, Zheng Zhedong, Wu Yu and Yang Yi, arXiv:1703.07220 [Attribute Dataset] |
| ACRN | 72.58% | 51.96% | "Person Re-Identification by Deep Learning Attribute-Complementary Information", Arne Schumann and Rainer Stiefelhagen, CVPR 2017 Workshop |
| PAN | 71.59% | 51.51% | "Pedestrian Alignment Network for Large-scale Person Re-identification", Zheng Zhedong, Zheng Liang and Yang Yi, arXiv:1707.00408 [code] |
| PAN+rerank | 75.94% | 66.74% | |
| SVDNet | 76.7% | 56.8% | "SVDNet for Pedestrian Retrieval", Sun Yifan, Zheng Liang, Deng Weijian and Wang Shengjin, ICCV 2017 |
| DPFL | 79.2% | 60.6% | "Person Re-Identification by Deep Learning Multi-Scale Representations", Chen Yanbei, Zhu Xiatian and Gong Shaogang, ICCV2017 workshop |
| SVDNet + REDA | 79.31% | 62.44% | "Random Erasing Data Augmentation", Zhong Zhun, Zheng Liang, Kang Guoliang, Shaozi Li and Yi Yang, arXiv:1708.04896, 2017. |
| SVDNet + REDA + ReRank | 84.02% | 78.28% |
We release our baseline training code and pretrained model in https://github.com/layumi/DukeMTMC-reID_baseline. Or you can directly download the finetuned ResNet-50 baseline feature. You can download it from GoogleDriver or BaiduYun, which includes the feature of training set, query set and gallery set. The DukeMTMC-reID LICENSE is also included.
We also annotated 23 human-level attributes (gender/clothing/...) for DukeMTMC-reID. You can find it in the following link: https://github.com/vana77/DukeMTMC-attribute
DukeMTMC Dataset [Bibtex]
DukeMTMC-reID Dataset, Protocol, Baseline [Bibtex]