The THU-MultiLiCa Dataset

Hardware Designs

Acquisition System

The system image of nvidai xavier will be available soon.

Manufacturing Designs

The assembly body is meant to hold the LiDAR and the camera in place so that the extrinsic calibration results can be reused, which can be readily done by manufacturing factories using open source manufacturing designs.
Download assembly body

Data orgnization

The data is organized as the following readme for each scene.

Datasets

Crossroads Scene

We collected 3D dynamic data of traffic scenes by 4 slave nodes set up at the four corners of a street crossroad, where many vehicles and pedestrians moved across the scene. We provided 4, 000 frames with 3D bounding box annotations of cars, cyclists, and pedestrians in the central area of the crossroad.
Readme for crossroads scene
Crossroads scene

Multi-person Scene

To evaluate the performance of long-term objects tracking, we collected data of multiple persons moving for a long time in the plaza and rarely leaving, for 1200 frames in the day and 1200 frames at night. The persons are free to move and interact with others, which results in a lot of occlusions and intersected trajectories, causing great difficulties for single-view data.
Readme for multi_person_day
Multi person scene for day

Readme for multi_person_night
Multi person scene at night

Code of Baselines

The implementation of the baselines is available at our github repository. [link]

Reference

Research papers that used this database should cite the following paper:

Meng Zhang, Wenxuan Guo, Bohao Fan, Yifan Chen, Jianjiang Feng, and Jie Zhou, “A Flexible Multi-view Multi-modal Imaging System for Large-Scale Outdoor Scenes”, 3DV, 2022

Contact

Contact Dr. Jianjiang Feng for further information.