Research map / Detection and segmentation / Open-set and few-shot detection
GoodSAM: Bridging Domain and Capacity Gaps via Segment Anything Model for Distortion-aware Panoramic Semantic Segmentation
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.
Distills SAM's zero-shot instance masks and boundaries, plus a Segformer teacher assistant, into a compact student for label-free panoramic semantic segmentation.
Problem
Learning a compact panoramic (ERP) semantic segmentation student without labeled data, given that SAM gives instance masks but no semantic labels and has a large capacity gap to the student; ERP distortion further hurts performance.
Approach
A teacher assistant (Segformer-B5) supplies semantic labels; the Distortion-Aware Rectification module fuses SAM instance masks and TA predictions (CTCF block with area-threshold and Shannon-entropy label selection), enforces overlap consistency and SAM-guided boundary enhancement, and the Multi-level Knowledge Adaptation module transfers whole-image and window-level knowledge to the student.
Key results
- GoodSAM-S (25.4M params) reaches 60.56% mIoU on DensePASS, 3.75% above DATR-S.
- GoodSAM-M with 3.7M parameters reaches 55.93% mIoU on DensePASS, comparable to DATR-T (54.60%) and above Trans4PASS-T (53.18%).
- On WildPASS, GoodSAM-S reaches 65.18% mIoU, 1.65% above DPPASS-S.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GoodSAM-S instead of DATR-S | DensePASS semantic segmentation · DensePASS test | mIoU: 56.81 -> 60.56 | Helped |
| Use GoodSAM-T instead of DATR-T | DensePASS semantic segmentation · DensePASS test | mIoU: 54.60 -> 58.21 | Helped |
| Use GoodSAM-M instead of DATR-M | DensePASS semantic segmentation · DensePASS test | mIoU: 52.90 -> 55.93 | Helped |
| Use GoodSAM-T instead of DPPASS-T | WildPASS semantic segmentation · WildPASS validation | mIoU: 60.38 -> 62.42 | Helped |
| Use GoodSAM-S instead of DPPASS-S | WildPASS semantic segmentation · WildPASS validation | mIoU: 63.53 -> 65.18 | Helped |
| Add sliding window sampling (ablation) | GoodSAM student on DensePASS · DensePASS ablation | mIoU: 15.88 -> 15.88 | Helped |
| Add CTCF fusion of SAM masks and TA predictions (ablation) | GoodSAM ensemble logits on DensePASS · DensePASS ablation | mIoU: 53.86 -> 55.88 | Helped |
| Add the consistency constraint (ablation) | GoodSAM teacher assistant on DensePASS · DensePASS ablation | mIoU: 58.89 -> 60.07 | Helped |
| Add boundary enhancement (ablation) | GoodSAM teacher assistant on DensePASS · DensePASS ablation | mIoU: 60.07 -> 62.49 | Helped |
| Add whole-image TA supervision (ablation) | GoodSAM student on DensePASS · DensePASS ablation | mIoU: 50.90 -> 54.12 | Helped |
| Add student boundary supervision (ablation) | GoodSAM student on DensePASS · DensePASS ablation | mIoU: 54.12 -> 55.93 | Helped |
| Replace CTCF with SSA fusion (ablation) | GoodSAM ensemble logits on DensePASS · DensePASS ablation | mIoU: 49.7 -> 47.93 | Hurt |
Surprising findings
- Segformer-B5 applied directly to ERP images without sliding windows drops to 27.62% mIoU, versus 53.86% with sliding windows.
Limitations
- SAM is limited in semantic information and has prohibitive parameters and domain gap, requiring a TA model.
- Requires sliding-window training with multiple windows per image, raising training cost as stride decreases.
Experimental setup
Datasets: WildPASS DensePASS
Benchmarks: DensePASS WildPASS
Metrics: mIoU
Baselines: ERFNet PASS Omni-sup P2PDA PCS Trans4PASS DPPASS DATR
Builds on: Segment Anything Model Trans4PASS DPPASS DATR Segformer
Closest papers by meaning
- Semantics Distortion and Style Matter: Towards Source-free UDA for Panoramic Segmentation (CVPR 2024 · similarity 0.92)
- Geometric Exploitation for Indoor Panoramic Semantic Segmentation (NeurIPS 2024 · similarity 0.86)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.85)
- SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.85)
- AdapToPASS: Ambiguity-aware Adaptive Spherical Transformer for Panoramic Semantic Segmentation (NeurIPS 2026 · similarity 0.85)
- Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.85)
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.