Person re-identification dataset summary

CUHK03

CUHK03 is the first person re-identification dataset large enough for deep learning. It provides bounding boxes and manual labeling detected from deformable part models (DPM). For this dataset, the quality of person detection is relatively good.

Test protocol:

There are two testing protocols for CUHK-03.

The first is the old version (reference [1], the source of the dataset), see the 'testsets' test protocol in the dataset. Specifically, 100 pedestrians are randomly selected as the test set, 1160 pedestrians as the training set, and 100 pedestrians as the verification set (there are a total of 1360 pedestrians instead of 1467, because the camera group pair is not used in the experiment. 4 and 5), repeated twenty times. This testing protocol is single-shot setting.

The second testing protocol (Reference [2]) is similar to Market-1501, which splits the dataset into a training set of 767 pedestrians and a test set of 700 pedestrians. In the testing phase, we randomly select one image as the query and the rest as the gallery, so that for each pedestrian, there are multiple ground truths in the gallery. (The new test protocol can refer to here)

link

Google Drive:

https://drive.google.com/file/d/0BxJeH3p7Ln48djNVVVJtUXh6bXc/edit?usp=sharing

Market1501

In 2015, the paper Person Re-Identification Meets Image Search proposed the Market 1501 dataset, and now the Market 1501 dataset has become one of the most commonly used datasets in the field of pedestrian re-identification.

Common disadvantages in databases are:

Small database size (few pictures)
The number of cameras is small (usually two)
fewer pedestrians
There is only one query for each identity
The pictures are hand-labeled picture perfect and lack practicality
In response to the above problems, Market1501 was established:

1501 identities
6 cameras
32668 images
DPM detector instead of manually framed pedestrians
500K interference pictures
Each identity has multiple queries
Each query corresponds to an average of 14.8 galleries
The pedestrian pictures of Market 1501 are collected from 6 cameras on the campus of Tsinghua University, and a total of 1501 pedestrians are marked. Among them, 751 pedestrian annotations are used for the training set, and 750 pedestrian annotations are used for the test set. There are no duplicate pedestrian IDs in the training set and the test set, which means that none of the 751 pedestrians that appear in the training set appear in the test set.

Training set: 751 pedestrians, 12936 images
Test set: 750 pedestrians, 19732 images
Query set: 750 pedestrians, 3368 images. The pedestrian pictures in the query set are all manually labeled pictures, and a picture is selected from 6 cameras for each pedestrian in the test set to form the query set. Each pedestrian in the test set has at most 6 images, and the query set has a total of 3368 images.
When the network model is trained, the training set will be used; when the model is tested, the test set and query set will be used. At this point the test set is also called the gallery set. Therefore, the subsets actually used are training set, gallery set and query set.

link

链接：百度网盘 请输入提取码 提取码：xdej

DukeMTMC-reID/DukeMTMC4ReID

link

链接: 百度网盘 请输入提取码 提取码: k3s