KITTI-360 - A Novel Dataset and Benchmarks for Urban Scene Understanding in 2D and 3D

This talk introduces KITTI-360, an urban driving dataset combining multiple sensors, comprehensive semantic and instance annotations, accurate localization, and benchmarks spanning computer vision, graphics, and robotics. It covers 3D-to-2D label transfer, annotation tools, semantic scene understanding, novel view synthesis, and semantic SLAM.

Syllabus

Intro
Combining Perception and Action
Combining Perception and Simulation
Towards Full Autonomy
Data Collection Data
The Curse of Dataset Annotation
3D to 2D Semantic and Instance Label Transfer
Data Annotation
Static Object Annotation
Semi-Automatic Dynamic Object Annotation
3D-to-2D Label Transfer
Qualitative Comparison to Baselines
Qualitative Results
Semantic Scene Understanding
Novel View Synthesis
Benchmarks
Semantic SLAM
Leaderboard
Resources