HeightFormer: Explicit Height Modeling Without Extra Data for Camera-Only 3D Object Detection in Bird's Eye View.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving Systems

Background:

• Vision-based Bird's Eye View (BEV) representation is crucial for autonomous driving perception.
• Generating accurate BEV representations from multi-camera systems is challenging due to its ill-posed nature.
• Existing methods often implicitly model depth or height, limiting their applicability.

Purpose of the Study:

• To propose an explicit height modeling approach for BEV representation generation.
• To develop a novel method, HeightFormer, that accurately estimates heights and uncertainties in BEV space.
• To demonstrate the effectiveness of camera-only BEV generation without requiring additional sensor data like LiDAR.

Main Methods:

• Explicitly modeling heights in the Bird's Eye View (BEV) space.
• Developing HeightFormer, a self-recursive model for height and uncertainty estimation.
• Proving the theoretical equivalence between height-based and depth-based BEV generation methods.

Main Results:

• HeightFormer accurately estimates heights in BEV without requiring extra data (e.g., LiDAR).
• The proposed method is adaptable to various camera configurations.
• HeightFormer achieves state-of-the-art performance compared to existing camera-only methods on benchmarks.

Conclusions:

• Explicitly modeling heights offers advantages for BEV generation in autonomous driving.
• HeightFormer provides a robust and accurate camera-only solution for BEV perception.
• The method advances the field of autonomous driving perception by improving BEV representation generation.