Alice Benchmarks
Alice is an online evaluation server that contains a series of domain adaptation tasks in computer vision. The source is simulated data; the target is real or synthetic data.
project page domain adaptive pedestrian recognition
More tasks are coming, aiming at content-level domain adaptation. That is, the source domain is editable through graphics engines.
PersonX engine
This is a data synthesis engine capable of generating persons under controllable
environmental conditions. That is, the illumination, viewpoint, camera settings can be edited precisely. In our CVPR19 paper, it is used to assess the impact of environmental factors on the system accuracy.
project page
CUHK03 dataset with new evaluation protocol
We have used a new evaluation protocol for the CUHK03 dataset. This protocol only has one train/test split instead of the previous 20 train/test splits. The training and testing sets have 767 and 700 IDs, respectively.
project page
MARS dataset
This dataset is a video extension of Market-1501. It has 6 cameras and 1,261 IDs. Only one train/test split is used.
project page state of the art on MARS
PRW dataset
This dataset is another extension of Market-1501. It provides 11,816 video frames. We draw a bounding box for all the pedestrians. We assign an ID (1-932) to most of them, and -2 to those whose IDs we do not know for sure. For this dataset, one should create a gallery by pedestrian detection and then evaluate person re-identification. Re-ID performance is closely related to detection performance.
project page