Autonomous Driving AI Challenge 2024 — Vehicle Object & State Recognition

1st of 146 teams · Director's Award · IITP (Korea's national ICT R&D agency)
Team lead · Nov 2024

▲ Qualitative results — per-vehicle instance masks across clear, hazy and low-visibility scenes.

Task

Recognize each vehicle and its composite state — class, lane position, and taillight action (multiple can hold at once) — with instance masks.

Model

A shared backbone and pixel decoder feed two heads: the transformer decoder emits class and mask, a parallel complex decoder reuses that mask for lane position and taillight state. Built on Mask2Former + DINOv2 with masked attention (baseline: YOLOv8). mAP 0.73 → 0.75.

Slides

13 slides (Korean)  ·  Open the PDF

Links

Ministry of Science and ICT press release (Korean, 19 Nov 2024) — lists our team as VIP (이정윤) in the vehicle composite-state recognition track.