Research map / Continual learning / Domain and prompt-based
Unsupervised Continual Semantic Adaptation Through Neural Rendering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Jointly train NeRF and DeepLab rather than fine-tuning only with rendered pseudo-labels (ablation) | DeepLabv3 ResNet-101 · ScanNet scenes 1–10, one-step validation views (Average) | mIoU: 45.9{ $\pm$1.4} -> 47.4{ $\pm$2.1} | Helped |
| Use NeRF-rendered images instead of ground-truth images for fine-tuning with NeRF pseudo-labels (ablation) | DeepLabv3 ResNet-101 · ScanNet scenes 1–10, one-step validation views (Average) | mIoU: 45.2{ $\pm$1.5} -> 45.9{ $\pm$1.4} | Helped |
| Use NeRF pseudo-labels instead of voxel-map pseudo-labels | DeepLabv3 ResNet-101 · ScanNet scenes 1–10, pseudo-labels on training views (Average) | mIoU: 44.0 -> 44.7{ $\pm$0.7} | Helped |
| Add depth supervision to NeRF training (ablation) | Semantic-NeRF pseudo-label generation · ScanNet pseudo-labels, scenes 1–10 | mIoU: +0.9% mIoU | Helped |
| Replace the original semantic loss with the modified semantic loss (ablation) | Semantic-NeRF pseudo-label generation · ScanNet pseudo-labels, scenes 1–10 | mIoU: +0.8% mIoU | Helped |
| Combine depth supervision with the modified semantic loss (ablation) | Semantic-NeRF pseudo-label generation · ScanNet pseudo-labels, scenes 1–10 | mIoU: 2.4% mIoU | Helped |
| Replay previous-scene NeRF renderings in addition to pre-training data (ablation) | Continual adaptation with DeepLabv3 ResNet-101 · ScanNet scenes 1–10, multi-step previous-scene retention (Average Prev.) | mIoU: 43.6{ $\pm$0.8} -> 44.6{ $\pm$0.6} | Helped |
| Replay NeRF renderings from novel viewpoints rather than training viewpoints (ablation) | Continual adaptation with DeepLabv3 ResNet-101 · ScanNet scenes 1–10, multi-step previous-scene retention (Average Prev.) | mIoU: 44.6{ $\pm$0.6} -> 45.3{ $\pm$0.9} | Helped |
| Jointly train NeRF and DeepLab rather than fine-tuning only (ablation) | DeepLabv3 ResNet-101 · ScanNet scenes 1–10, one-step validation views | mIoU: Scene 5: 52.8{ $\pm$2.9} -> 42.7{ $\pm$1.0}; Average: 45.9{ $\pm$1.4} -> 47.4{ $\pm$2.1} | Mixed |
| Use original Semantic-NeRF rather than Instant-NGP (ablation) | Semantic-NeRF pseudo-label generation · ScanNet Scene 1 training views | mIoU: 48.4{ $\pm$0.9} -> 50.0 { $\pm$0.5} | No clear effect |
Closest papers by meaning
- Unsupervised Model Adaptation for Continual Semantic Segmentation (AAAI 2021 · similarity 0.91)
- On the Road to Online Adaptation for Semantic Image Segmentation (CVPR 2022 · similarity 0.89)
- GP-NeRF: Generalized Perception NeRF for Context-Aware 3D Scene Understanding (CVPR 2024 · similarity 0.89)
- Towards Continual Universal Segmentation (CVPR 2025 · similarity 0.87)
- Pay Attention to Target: Relation-Aware Temporal Consistency for Domain Adaptive Video Semantic Segmentation (AAAI 2024 · similarity 0.87)
- UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior (NeurIPS 2024 · 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.