Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with Self-Distillation

Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with Self-Distillation

18 Apr 2024 | Song Wang¹, Jiawei Yu¹, Wentong Li¹, Wenyu Liu¹, Xiaolu Liu¹, Junbo Chen², Jianke Zhu¹*
This paper proposes a hardness-aware semantic scene completion (HASSC) approach to improve the accuracy of semantic scene completion models without incurring extra inference costs. The HASSC method incorporates global and local hardness to identify and refine challenging voxels during training. Global hardness is determined based on the uncertainty in predicting each voxel, while local hardness is derived from geometric anisotropy, measuring the semantic difference between a voxel and its neighbors. A self-distillation strategy is introduced to enhance model stability and consistency. The HASSC approach is integrated into existing models and has been validated through extensive experiments on the SemanticKITTI dataset. The results show that HASSC significantly improves the performance of baseline models, particularly in challenging regions. The method is efficient, with minimal overhead during training and inference. The HASSC approach outperforms existing methods in semantic scene completion, demonstrating its effectiveness in improving the accuracy of autonomous driving systems. The code is available at https://github.com/songw-zju/HASSC.This paper proposes a hardness-aware semantic scene completion (HASSC) approach to improve the accuracy of semantic scene completion models without incurring extra inference costs. The HASSC method incorporates global and local hardness to identify and refine challenging voxels during training. Global hardness is determined based on the uncertainty in predicting each voxel, while local hardness is derived from geometric anisotropy, measuring the semantic difference between a voxel and its neighbors. A self-distillation strategy is introduced to enhance model stability and consistency. The HASSC approach is integrated into existing models and has been validated through extensive experiments on the SemanticKITTI dataset. The results show that HASSC significantly improves the performance of baseline models, particularly in challenging regions. The method is efficient, with minimal overhead during training and inference. The HASSC approach outperforms existing methods in semantic scene completion, demonstrating its effectiveness in improving the accuracy of autonomous driving systems. The code is available at https://github.com/songw-zju/HASSC.
Reach us at info@study.space