🚗 OpenDriveFM — Trust-Aware BEV Perception [LIVE]

Camera-only BEV occupancy · GPT-2 trajectory · Self-supervised fault detection · 317 FPS · ADE=2.457m

Real nuScenes data · Real model inference · Apple Silicon MPS · LIU Image and Vision Computing

🎛️ Controls

0 81
⚖️ Fusion Mode [T / W / U / R]

🎯 Fault Injection — Per Camera

UNSEEN = not in training set (generalization test)

CAM1 FRONT
CAM2 F-LEFT
CAM3 F-RIGHT
CAM4 BACK
CAM5 B-LEFT
CAM6 B-RIGHT

⚡ Quick Inject All Cameras

🔧 Mode Buttons

🎬 Features

🌦️ Generalization (UNSEEN faults)

🔬 Sparse Attention Mode

V — Sparse Mode

📽️ BEV Forecast Frame

G — Forecast Frame

📋 Key Numbers

Metric Value
FPS 317 (MPS)
Latency p50 3.15 ms
C++ Latency 4.449 ms
Traj ADE 2.457 m
BEV IoU 0.136
Parameters 553K
Trust detection 100%

🗺️ BEV Occupancy Map

📊 Trust Scores

📷 All 6 Cameras — Real nuScenes Data

🔬 Ablation Study

OpenDriveFM · LIU Image and Vision Computing · April 2026 · Akila Lourdes · Akilan Manivannan GitHub