The paper "Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with Self-Distillation" addresses the challenge of semantic scene completion (SSC) in autonomous driving, focusing on the issue that existing methods often treat all voxels equally, neglecting the varying hardness of voxels. The authors propose a hardness-aware semantic scene completion (HASSC) approach that integrates hard voxel mining (HVM) and self-distillation to improve the model's performance. **Key Contributions:** 1. **HASSC Approach:** HASSC is designed to train a semantic scene completion model with hardness-aware design, selecting and refining hard voxels dynamically. 2. **Global and Local Hardness:** Global hardness is defined based on the uncertainty in predicting each voxel, while local hardness is based on geometric anisotropy, which measures the semantic difference from neighbors. 3. **Self-Distillation:** A self-distillation strategy is introduced to stabilize and enhance the training process by using the teacher model's exponential moving average (EMA) to update the student model. **Methodology:** - **HVM Head:** The HVM head selects hard voxels based on global hardness and refines them using local hardness. It dynamically updates the selected hard voxels during training. - **Self-Distillation:** The teacher model, which shares the same architecture as the student model, is updated using the EMA of the student model's parameters. This ensures stable and consistent training. **Experiments:** - **SemanticKITTI Dataset:** Extensive experiments on the SemanticKITTI dataset demonstrate that HASSC significantly improves the accuracy of existing models without increasing inference costs. - **Ablation Studies:** Ablation experiments show that the combination of global and local hardness, self-distillation, and appropriate hyperparameters are crucial for the model's performance. - **Comparison with Baselines:** HASSC outperforms state-of-the-art camera-based methods, including VoxFormer-S and VoxFormer-T, on both quantitative and qualitative metrics. **Conclusion:** HASSC effectively addresses the issue of treating all voxels equally in semantic scene completion, leading to improved performance in challenging regions. The proposed approach is generic and can be easily integrated into existing models, making it a valuable contribution to the field of autonomous driving perception.The paper "Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with Self-Distillation" addresses the challenge of semantic scene completion (SSC) in autonomous driving, focusing on the issue that existing methods often treat all voxels equally, neglecting the varying hardness of voxels. The authors propose a hardness-aware semantic scene completion (HASSC) approach that integrates hard voxel mining (HVM) and self-distillation to improve the model's performance. **Key Contributions:** 1. **HASSC Approach:** HASSC is designed to train a semantic scene completion model with hardness-aware design, selecting and refining hard voxels dynamically. 2. **Global and Local Hardness:** Global hardness is defined based on the uncertainty in predicting each voxel, while local hardness is based on geometric anisotropy, which measures the semantic difference from neighbors. 3. **Self-Distillation:** A self-distillation strategy is introduced to stabilize and enhance the training process by using the teacher model's exponential moving average (EMA) to update the student model. **Methodology:** - **HVM Head:** The HVM head selects hard voxels based on global hardness and refines them using local hardness. It dynamically updates the selected hard voxels during training. - **Self-Distillation:** The teacher model, which shares the same architecture as the student model, is updated using the EMA of the student model's parameters. This ensures stable and consistent training. **Experiments:** - **SemanticKITTI Dataset:** Extensive experiments on the SemanticKITTI dataset demonstrate that HASSC significantly improves the accuracy of existing models without increasing inference costs. - **Ablation Studies:** Ablation experiments show that the combination of global and local hardness, self-distillation, and appropriate hyperparameters are crucial for the model's performance. - **Comparison with Baselines:** HASSC outperforms state-of-the-art camera-based methods, including VoxFormer-S and VoxFormer-T, on both quantitative and qualitative metrics. **Conclusion:** HASSC effectively addresses the issue of treating all voxels equally in semantic scene completion, leading to improved performance in challenging regions. The proposed approach is generic and can be easily integrated into existing models, making it a valuable contribution to the field of autonomous driving perception.