GSH-Occ: Gradient-Shielded and Height-Aware BEV Occupancy Network.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving

Background:

• Camera-based 3D occupancy prediction often uses bird's-eye-view (BEV) representations.
• Existing methods face challenges with optimization instability and preserving vertical scene structure during feature lifting.

Purpose of the Study:

• To introduce GSH-Occ, a novel framework designed to overcome limitations in camera-based 3D occupancy prediction.
• To improve the stability of integrating new modules into pre-trained BEV encoders.
• To enhance the preservation of elevation-aware scene structure in occupancy prediction.

Main Methods:

• GSH-Occ employs Gradient-Shielded Residual Dual Attention (GS-RDA) with a zero-initialized residual gate for stable module insertion.
• Height-Aware Adaptive Lift (HAL) uses per-voxel adaptive fusion of BEV features and height embeddings, followed by 3D convolutional refinement.
• The framework integrates these mechanisms to address distinct failure modes in occupancy prediction.

Main Results:

• GSH-Occ achieved 46.92 mIoU on the Occ3D-nuScenes validation benchmark.
• The proposed method outperformed FlashOcc by 3.40 mIoU.
• Ablation studies validated the complementary improvements from GS-RDA and HAL.

Conclusions:

• GSH-Occ effectively addresses optimization instability and height-agnostic limitations in BEV-based 3D occupancy prediction.
• The combination of GS-RDA and HAL leads to significant performance gains.
• The framework demonstrates a robust approach for accurate 3D scene understanding from camera data.