We propose to organize this Multi-camera Multiple People Tracking workshop, aiming to gather academic and industry community together to tackle indoor multiple people tracking using multiple RGB cameras.

Overview

Multiple object tracking is one of the most basic and most important tasks in computer vision. It is one of the fundamental research topics in understanding visual content. It has numerous applications in indoor navigation, motion capture, human computer interaction, robotics etc.

For single-camera tracking, there are several datasets and benchmarks, which stimulate novel ideas of tracking models using sequential information. For multi-camera tracking, data collection and labeling are much more difficult. Thus, there are only a few small datasets available and no common benchmarks.

To support the community to develop more efficient and novel tracking algorithms, we construct a multi-camera multiple people tracking dataset. We expect our datasets can also be served as an evaluation benchmark for this task. This challenge invites all research teams to participate, and dataset will be released to the research community. We invite academic and industrial researchers to participate in this multi-camera multiple people tracking contest.

Multi-camera Multiple People Tracking (MMP-Tracking) Challenge

MMP-Tracking Challenge aims to push the state of the art multi-camera multiple people tracking algorithms forward. The participants will receive an annotated training set and a test set without annotations.

There are two subtracks in this challenge. (1) Evaluating tracking results from topdown view; (2) Evaluating tracking results from each camera view, then aggregate together for final metric.

For subtrack (1), using camera calibration files provided, one can mapping the ground plane in each camera view into a world coordinate shared by all cameras. Then discretize the world coordinate with voxel size 20mm to get topdown view map. We provide ground truth label of person footpoint coordinates in topdown view map. During evaluation, we compute false positive(FP), false negative(FN) and true positive(TP) by assigning detected person footpoint to ground truth footpoint using Hungarian matching. We impose that a detected footpoint can be assigned to ground truth only if they are less than 0.5m away (25pixel distance in topdown map). After getting FP, FN and TP, we will compute IDF1 and MOTA as our final score.

For subtrack (2), following the same evaluation procedure widely used in MOT, we compute FP, FN and TP for each camera view independently (detected bounding box and ground truth bounding box can be matched only their IOU>0.5). After getting results from each camera view, we average over all these results to computer our final metrics. We will use IDF1 and MOTA as our final score.

Our evaluation codes are based on py-motmetrics . The evaluation codes can be downloaded here for local usage by participants. We use CodaLab as our evaluation server link .

Winners

• Track 1 - Topdown View:

Rank	Team
1st	Alibaba Group	Fei Du, Jiasheng Tang, et al.
2nd	Hikvision Research Institute	Jie Zhang, et al.
3rd	University of Science and Technology of China	Size Wu
4th	Fujitsu Research	Shoichi Masui, Fan Yang, et al.

• Track 2 - Camera View:

Rank	Team
1st	Hikvision Research Institute	Jie Zhang, et al.
2nd	Wyze Labs, Inc, AI Team	Hung-Min Hsu, et al.
3rd	Information Processing Lab, University of Washington	Cheng-Yen Yang, et al.