CMC Dataset - Multi-view/Multi-camera 3D People Tracking Dataset
The new CMC dataset is a 4-camera 1920x1024 resolution dataset recorded at 4fps in a room with dimensions 7.67m x 3.41m x 2.7m. The CMC dataset has 5 different sequences with varying levels of person density and occlusion.
Description:
- CMCX -> Folders containing download links (recommend using Chrome to download), i.e.,
- CMC1 has a maximum of 3 people and virtually no occlusion.
- CMC2 has a maximum of 10 people with some occlusion.
- CMC3 has a maximum of 15 people with significant occlusion.
- CMC4 involves people jumping and falling with a maximum of 3 people.
- CMC5 involves people jumping and falling with a maximum of 7 people.
- MOT format annotations
- Reference: https://motchallenge.net/instructions/
- GT_CMCX_Cam_X.txt -> Bounding-box ground truths for each camera.
- DET_CMCX_CamX.txt -> Bounding-box detections for each camera.
- GT_CMCX_WOLD_CENTROID.txt -> 3D centroid + extent (axis-aligned ellipsoids) ground truths, i.e., <x_c y_c z_c x_half y_half z_half>, where <x_c y_c z_c> is the centroid and <x_half y_half z_half> is the half-lengths of the ellipsoid.
- This 4-camera dataset is published in relation to the following paper:
-
@article{ong2020bayesian,
author={Ong, Jonah and Vo, Ba-Tuong and Vo, Ba-Ngu and Kim, Du Yong and Nordholm, Sven},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
title={A {B}ayesian Filter for Multi-View 3{D} Multi-Object Tracking With Occlusion Handling},
year={2022},
volume={44},
number={5},
pages={2246-2263},
doi={10.1109/TPAMI.2020.3034435.}
}
- Once you download the datasets, place all subfolders in a master folder as shown below.
