Semantic KITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences

Date: 2019
Journal: CVPR

1. Introduction

LiDAR sensors are not affected by lighting, providing precise distance measurements

SemanticKITTI focuses on laser based semantic segmentation and semantic scene completion

2. Related Work

3. The Semantic KITTI

3.1. Labeling Process

Loop close the sequences using an off the shelf laser based SLAM system

Subdivide the sequence of point clouds into tiles of 100m by 100m

For each tile, load scans overlapping with tile, enabling to label all scans consistently

3.2. Dataset Statistics

The unbalanced count of classes occured, but is common for data from natural environments

4. Evaluation of Semantic Segmentation

4.1. Single Scan Experiments

Task and Metrics

Used method, commonly applied mean Jaccard Index or mean intersection over union (mIOU) metric

Cannot expect to distinguish moving from non-moving objects with single scan

State of the Art

Feature extraction and classification is replaced by end to end deep neural networks (CNN) with 3D convolutions for object classification and semantic segmentation

To overcome the limitation of voxel based representation such as exploding memory consumption, recent approaches either upsample voxel predictions using CRF or use different representations

Baseline approaches