Research map / 3D vision / 3D generation
Memory-Augmented Re-Completion for 3D 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 MARE, a camera-based semantic scene completion model that uses a region-token memory bank with diversity/age update rules and a re-completion pipeline to fill invisible voxels.
Problem
Camera-based 3D semantic scene completion produces many false negatives in invisible regions (occluded or outside the field of view) because input information for those areas is missing.
Approach
Stores region-tokens from visible areas in a regional memory bank, updated by diversity and age scores, and iteratively re-completes invisible voxels by querying the bank, with an ASPP-based Information Spreading module bridging region-level and voxel-level scales.
Key results
- On SSCBench-KITTI-360 test, MARE reaches 18.84 mIoU and 57.09 recall, improving over Symphonies by 0.26 mIoU and 2.21 recall.
- On SemanticKITTI val, MARE reaches 15.39 mIoU and 65.38 recall, surpassing Symphonies (14.89 mIoU in table).
- Ablation shows the Regional Memory Bank adds about 0.24 mIoU over the re-completion pipeline alone (18.60 to 18.84).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MARE instead of Symphonies | Camera-based SSC · SSCBench-KITTI-360 test | IoU: 44.12 -> 46.10 (+2.02) | Helped |
| Use MARE instead of Symphonies | Camera-based SSC · SSCBench-KITTI-360 test | precision: 69.24 -> 70.53 (+1.31) | Helped |
| Use MARE instead of Symphonies | Camera-based SSC · SSCBench-KITTI-360 test | recall: 54.88 -> 57.09 (+2.21) | Helped |
| Use MARE instead of Symphonies | Camera-based SSC · SSCBench-KITTI-360 test | mIoU: 18.58 -> 18.84 (+0.26) | Helped |
| Use MARE instead of Symphonies | Camera-based SSC · SemanticKITTI val | recall: 61.16 -> 65.38 | Helped |
| Use MARE instead of Symphonies | Camera-based SSC · SemanticKITTI val | precision: 62.17 -> 56.01 | Hurt |
| Add the Re-completion pipeline to SSC Transformer (ablation) | MARE · SSCBench-KITTI-360 | IoU: 43.80 -> 45.99 | Helped |
| Add the Re-completion pipeline to SSC Transformer (ablation) | MARE · SSCBench-KITTI-360 | mIoU: 18.25 -> 18.60 | Helped |
| Add the Regional Memory Bank to Re-completion (ablation) | MARE · SSCBench-KITTI-360 | mIoU: 18.60 -> 18.84 | Helped |
| Add a second re-completion iteration (medium distance) (ablation) | MARE · SemanticKITTI | recall: 61.00 -> 65.38 | Helped |
| Add a third re-completion iteration (far distance) (ablation) | MARE · SemanticKITTI | mIoU: 15.39 -> 14.01 | Hurt |
| Add a third re-completion iteration (far distance) (ablation) | MARE · SemanticKITTI | precision: 56.01 -> 60.05 | Helped |
Surprising findings
- Precision drops on SemanticKITTI relative to Symphonies, which the authors attribute to label inconsistencies.
- Some categories such as truck and other-vehicle decline in IoU, which the authors attribute to the injection mechanism disturbing fine edge features.
Limitations
- Performance gains on SemanticKITTI are smaller, and precision declines there.
- Injection mechanism degrades fine-grained categories such as truck and other-vehicle.
- Uses only the current frame; no temporal information.
Experimental setup
Datasets: SSCBench-KITTI-360 SemanticKITTI
Benchmarks: SSCBench-KITTI-360 SemanticKITTI
Metrics: IoU precision recall mIoU
Baselines: MonoScene TPVFormer VoxFormer OccFormer NDC-Scene H2GFormer Symphonies
Builds on: Symphonies MonoScene MaskDINO VoxFormer
Closest papers by meaning
- Three Cars Approaching within 100m! Enhancing Distant Geometry by Tri-Axis Voxel Scanning for Camera-based Semantic Scene Completion (CVPR 2025 · similarity 0.90)
- Unleashing Network Potentials for Semantic Scene Completion (CVPR 2024 · similarity 0.90)
- Context and Geometry Aware Voxel Transformer for Semantic Scene Completion (NeurIPS 2024 · similarity 0.90)
- 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)
- VLScene: Vision-Language Guidance Distillation for Camera-Based 3D Semantic Scene Completion (AAAI 2025 · similarity 0.89)
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.