CoBEV: Elevating Roadside 3D Object Detection with Depth and Height Complementarity.

Area of Science:

• Computer Vision
• Intelligent Transportation Systems
• 3D Object Detection

Background:

• Roadside camera-based 3D object detection is vital for intelligent transportation systems, improving safety beyond vehicle-centric perception.
• Existing methods often rely on limited depth or height information, overlooking their complementary nature.

Purpose of the Study:

• To develop a novel framework, Complementary-BEV (CoBEV), for end-to-end monocular 3D object detection.
• To leverage the complementary geometric and semantic information from both depth and height features for robust Bird's-Eye View (BEV) representations.

Main Methods:

• CoBEV estimates pixel-wise depth and height distributions to lift camera features into 3D space.
• A two-stage complementary feature selection (CFS) module facilitates lateral fusion.
• A BEV feature distillation framework is integrated, utilizing a CoBEV teacher model for enhanced accuracy.

Main Results:

• CoBEV achieves state-of-the-art accuracy on public benchmarks (DAIR-V2X-I, Rope3D) and a private dataset (Supremind-Road).
• Demonstrates significant improvements in robustness against long-distance challenges and camera noise.
• Achieves a 80% vehicle AP score on DAIR-V2X-I (easy mode) for the first time with a camera-only model.

Conclusions:

• CoBEV effectively integrates complementary depth and height information for superior 3D object detection.
• The framework significantly enhances robustness and generalization capabilities in diverse and challenging roadside scenarios.
• CoBEV represents a substantial advancement in camera-based 3D object detection for intelligent transportation systems.