Pose Estimation

Given the background data and an object instance, the pose estimation module aims to estimate possible valid locations and orientations to create a plausible scene in the real world after inserting the object.

Here we provide more visualisations of pose estimation module in MulTest, which we were unable to provide due to the limited space of the paper.

Pose Generation

This module aims to find a suitable insertion position in the 3D scene

First, MultiTest utilize CENet [1], a concise and efficient point cloud semantic segmentation model, to split the road point cloud from the background point cloud. Compared to plane-equation based sampling method (such as RANSAC), it can get more accurate road segmentation even in complicated road conditions such as intersections. Here are some examples.

In contrast, MulTest's road segmentation module (1) is able to segment different types of roads, and (2) does not treat non-road planes (e.g., sidewalk) as roads.

Segmentation results under different types of roads

As the road point cloud is denser around the LiDAR, sampling the positions of the inserted objects directly on the segmented road may could lead to limited and uneven distribution of the positions. To alleviate it, we meshed the road point cloud to reconstruct the pavement and obtain uniformly distributed sample locations for object insertion. Here are some examples.

Collision Avoidance

This module aims to avoid collisions and check if the pose generated by the pose generator is valid.

Inserting objects with a given position may conflict or collide with the background or existing objects. Here are some examples.

Reference

[1] Hui-Xian Cheng, Xian-Feng Han, and Guo-Qiang Xiao. 2022. Cenet: Toward concise and efficient lidar semantic segmentation for autonomous driving. In 2022 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 01–06

Google Sites

Report abuse