Research map / 3D vision / 3D generation
Three Cars Approaching within 100m! Enhancing Distant Geometry by Tri-Axis Voxel Scanning for Camera-based Semantic Scene Completion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ScanSSC, a camera-based semantic scene completion model with axis-wise near-to-far masked attention (Scan Module) and cumulative-average cross-entropy (Scan Loss) to improve distant geometry.
Problem
Camera-based semantic scene completion underestimates geometry in distant regions due to perspective and occlusion, which is critical for driving safety.
Approach
Scan each voxel axis near-to-far with cascaded masked self-attention so distant voxels attend to nearer voxels, and add a loss that cross-entropies cumulatively averaged logits so context propagates to distant voxels.
Key results
- ScanSSC reaches IoU 44.54 and mIoU 17.40 on SemanticKITTI hidden test, beating CGFormer's 44.41 and 16.63.
- On SSCBench-KITTI-360 test, ScanSSC achieves IoU 48.29 and mIoU 20.14, slightly above CGFormer (48.07, 20.05).
- Ablation shows the Scan Module and Scan Loss together raise IoU by 3.72 and mIoU by 2.21 over baseline on SemanticKITTI validation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ScanSSC instead of CGFormer | camera-based SSC · SemanticKITTI hidden test set | IoU: 44.41 -> 44.54 | Helped |
| use ScanSSC instead of HTCL-S | camera-based SSC · SemanticKITTI hidden test set | mIoU: 17.09 -> 17.40 | Helped |
| use ScanSSC instead of CGFormer | camera-based SSC · SSCBench-KITTI-360 test set | IoU: 48.07 -> 48.29 | Helped |
| use ScanSSC instead of CGFormer | camera-based SSC · SSCBench-KITTI-360 test set | mIoU: 20.05 -> 20.14 | Helped |
| add Scan Module and Scan Loss (ablation) | SemanticKITTI SSC baseline · SemanticKITTI validation set | mIoU: 14.91 -> 17.12 | Helped |
| add depth-, width-, and height-axis Scan Module blocks (ablation) | SemanticKITTI SSC baseline · SemanticKITTI validation set | mIoU: 14.91 -> 16.63 | Helped |
| add depth-, width-, and height-axis Scan Loss (ablation) | SemanticKITTI SSC baseline · SemanticKITTI validation set | mIoU: 14.91 -> 15.19 | Helped |
| reverse depth-axis refinement direction in the Scan Module (ablation) | ScanSSC · SemanticKITTI validation set | mIoU: 17.12 -> 16.73 | Hurt |
| reverse width-axis refinement direction in the Scan Module (ablation) | ScanSSC · SemanticKITTI validation set | mIoU: 17.12 -> 16.99 | Hurt |
| reverse height-axis refinement direction in the Scan Module (ablation) | ScanSSC · SemanticKITTI validation set | mIoU: 17.12 -> 16.96 | Hurt |
| reverse depth-axis direction in Scan Loss (ablation) | ScanSSC · SemanticKITTI validation set | mIoU: 17.12 -> 16.51 | Hurt |
| reverse width-axis direction in Scan Loss (ablation) | ScanSSC · SemanticKITTI validation set | mIoU: 17.12 -> 16.09 | Hurt |
Surprising findings
- Scan Loss alone has a relatively small effect (+0.85 IoU, +0.28 mIoU), while the Scan Module contributes most of the gain.
Limitations
- Mask margin ratios are hand-set per axis and sensitive to deviation from defaults.
- Gains over CGFormer on SSCBench-KITTI-360 are small (about 0.2 IoU and 0.1 mIoU).
Experimental setup
Datasets: SemanticKITTI SSCBench-KITTI-360
Benchmarks: SemanticKITTI SSCBench-KITTI-360
Metrics: IoU mIoU recall axis-wise mIoU
Baselines: VoxFormer-T CGFormer HTCL-S Symphonies MonoScene TPVFormer OccFormer SurroundOcc IAMSSC H2GFormer-T HASSC-T StereoScene
Builds on: CGFormer VoxFormer MonoScene LSS masked self-attention (Transformer)
Closest papers by meaning
- Memory-Augmented Re-Completion for 3D Semantic Scene Completion (AAAI 2025 · similarity 0.90)
- Context and Geometry Aware Voxel Transformer for Semantic Scene Completion (NeurIPS 2024 · similarity 0.89)
- Learning Temporal 3D Semantic Scene Completion via Optical Flow Guidance (NeurIPS 2025 · similarity 0.89)
- Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with Self-Distillation (CVPR 2024 · similarity 0.89)
- H2GFormer: Horizontal-to-Global Voxel Transformer for 3D Semantic Scene Completion (AAAI 2024 · similarity 0.89)
- VLScene: Vision-Language Guidance Distillation for Camera-Based 3D Semantic Scene Completion (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.