HeLiMOS: Dataset for Heterogeneous LiDAR Moving Object Segmentation

Why HeLiMOS?

We build upon the existing HeLiPR dataset and provide MOS labels that enable the evaluation of MOS across four heterogeneous LiDAR sensor setups, which we call HeLiMOS.

Our dataset is based on the KAIST05 sequence of HeLiPR dataset, which contains various moving objects, such as buses, pedestrians, bicyclists, and cars, different from other sequences.

The dataset is acquired by four LiDAR sensors: Velodyne VLP-16 and Ouster OS2-128 as omnidirectional LiDAR sensors, and Livox Avia and Aeva Aeries II as solid-state LiDAR sensors, which have irregular scanning patterns compared with the traditional spinning LiDAR sensors.

Folder structure and format

Our dataset provides a total of 12,188 labeled point clouds. Each MOS label follows the SemanticKITTI-MOS format, so it consists of three classes: unlabeled, static, and dynamic.

Novel characteristics of the dataset

1. A variety of dynamic points patterns

Because we exploited the HeLiPR dataset, whose scans are captured repeatedly in the same urban locations to evaluate place recognition tasks, the MOS labels of our HeLiMOS show a variety of dynamic point patterns owing to the varying trajectories of moving objects even though the scans are acquired in the same places.

2. Significantly higher ratio of dynamic points

The most distinctive feature of our dataset is that it not only has higher dynamic points ratios than existing MOS datasets, but also has MOS labels of four heterogeneous LiDAR sensors.

Our dataset shows consistently higher average dynamic points ratios across all LiDAR sensors compared with the SemanticKITTI and SemanticPOSS datasets.

3. Point-wise MOS labels captured from different LiDAR sensors

(a) Aeva Aeries II

(b) Livox Avia

(d) Velodyne VLP-16

What's next?

Page updated

Google Sites

Report abuse