Research map / Detection and segmentation / Open-set and few-shot detection
TALoS: Enhancing Semantic Scene Completion via Test-time Adaptation on the Line of Sight
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 TALoS, a test-time adaptation method for LiDAR semantic scene completion that uses line-of-sight emptiness and cross-moment observations as self-supervision, with a dual delayed-update scheme for future frames.
Problem
LiDAR semantic scene completion is ill-posed and relies on training-set structural priors, so pre-trained SSC models generalize poorly to diverse driving scenes.
Approach
Use observations at one moment as pseudo ground truth for another: occupied voxels from projected points plus empty voxels along the LiDAR line of sight (Bresenham), aggregated reliable semantic pseudo-labels across moments, and a delayed update of a gradual model that incorporates future frames once they arrive.
Key results
- On SemanticKITTI validation, TALoS raises SCPNet from 37.56 to 39.29 mIoU and from 50.24 to 56.09 cIoU.
- On the SemanticKITTI test set (online server), TALoS reaches 37.9 mIoU and 60.2 cIoU versus SCPNet's 36.7 and 56.1.
- Cross-dataset SemanticKITTI-to-SemanticPOSS yields cIoU of 36.2 vs 25.2 for the SCPNet baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply TALoS test-time adaptation | SCPNet · SemanticKITTI test | mIoU: 36.7 -> 37.9 | Helped |
| apply TALoS test-time adaptation | SCPNet · SemanticKITTI test | cIoU: 56.1 -> 60.2 | Helped |
| apply TALoS test-time adaptation | SCPNet · SemanticPOSS val | cIoU: 25.2 -> 36.2 | Helped |
| add completion loss to moment-wise adaptation (ablation) | SCPNet on SemanticKITTI val · SemanticKITTI val | cIoU: 50.24 -> 52.81 | Helped |
| add semantic loss to moment-wise adaptation (ablation) | SCPNet on SemanticKITTI val · SemanticKITTI val | mIoU: 37.56 -> 38.35 | Helped |
| add both completion and semantic losses to moment-wise adaptation (ablation) | SCPNet · SemanticKITTI val | mIoU: 37.56 -> 38.38 | Helped |
| add gradual adaptation to moment-wise adaptation (ablation) | TALoS · SemanticKITTI val | mIoU: 38.38 -> 39.29 | Helped |
| add moment-wise adaptation to gradual adaptation (ablation) | TALoS · SemanticKITTI val | cIoU: 55.94 -> 56.09 | Helped |
| replace TALoS losses with entropy minimization | SCPNet with moment-wise adaptation · SemanticKITTI val | cIoU: 50.24 -> 49.86 | Hurt |
| use CoTTA entropy-minimization adaptation | SCPNet · SemanticKITTI val | mIoU: 37.56 -> 36.55 | Hurt |
| increase the number of moment-model update iterations (ablation) | TALoS · SemanticKITTI val | mIoU | Helped |
| increase calibration-transform noise (ablation) | TALoS · SemanticKITTI val | mIoU | Hurt |
Surprising findings
- Entropy-minimization TENT slightly lowers cIoU below the baseline (49.86 vs 50.24), and CoTTA reduces mIoU (36.55 vs 37.56).
- Naive early and late temporal fusion reduce mIoU (35.86 and 37.11 vs 37.56) even though they raise cIoU.
Limitations
- Only addresses point clouds; image-based SSC not covered.
- Gains on the severe domain gap (SemanticPOSS) remain low in mIoU because the pre-trained model is weak.
- Adds roughly 3x per-step computation over the baseline and depends on LiDAR calibration accuracy.
Experimental setup
Datasets: SemanticKITTI SemanticPOSS
Benchmarks: SemanticKITTI SemanticPOSS KITTI-360
Metrics: mIoU cIoU
Baselines: SCPNet TENT CoTTA SSA-SC JS3C-Net S3CNet SSCNet early fusion late fusion
Builds on: SCPNet TENT CoTTA Lovasz-softmax loss
Closest papers by meaning
- Towards Learning to Complete Anything in Lidar (ICML 2025 · similarity 0.88)
- Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation (CVPR 2025 · similarity 0.87)
- Is Your LiDAR Placement Optimized for 3D Scene Understanding? (NeurIPS 2024 · similarity 0.86)
- IPFormer: Visual 3D Panoptic Scene Completion with Context-Adaptive Instance Proposals (NeurIPS 2025 · similarity 0.86)
- HiLoTs: High-Low Temporal Sensitive Representation Learning for Semi-Supervised LiDAR Segmentation in Autonomous Driving (CVPR 2025 · similarity 0.86)
- VLScene: Vision-Language Guidance Distillation for Camera-Based 3D Semantic Scene Completion (AAAI 2025 · similarity 0.86)
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.