1. Speaker:
Sanmin Kim (Department of Automotive IT Convergence, Kookmin University)

2. Title:
Vision-based 3D Perception for Autonomous Driving

3. Date & Time:
Thursday, May 28, 2026, 4:00 PM–5:30 PM

4. Venue:
L409

5. Abstract:
Cameras offer rich semantic cues and cost-effective scalability, making vision-based 3D perception a central pillar of modern autonomous driving. This talk briefly reviews the evolution of image-based 3D object detection along spatial and temporal dimensions, together with training strategies that mitigate the inherent depth ambiguity of cameras and 3D occupancy prediction as a complementary scene representation.

Building on this foundation, the talk discusses how the shift toward end-to-end autonomous driving requires a fundamental rethinking of perception itself. Rather than functioning as an isolated module, perception should become planning-aware and reactive by integrating feedback from prediction and planning to dynamically adapt perceptual saliency according to driving contexts. The talk will also address the importance of planning-centric evaluation metrics that directly reflect downstream autonomous driving performance.