Research map / Detection and segmentation / Open-set and few-shot detection
Style Evolving along Chain-of-Thought for Unknown-Domain Object Detection
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 Style Evolving along Chain-of-Thought, which builds progressively richer CLIP text prompts (words to phrases to sentences) to drive AdaIN-style feature evolution plus style disentanglement and class-specific prototypes, for single-domain generalized object detection.
Problem
Single-domain generalized object detection (Single-DGOD) must generalize to unseen target domains without target data; one-step textual prompts struggle to capture combined styles such as rain plus night.
Approach
Replace single-step text prompts with a chain of thought that progresses from simple to complex and coarse to fine, using CLIP text features to learn AdaIN-style parameters that evolve source-domain features; decouple style from content and add class-specific prototypes to preserve semantics.
Key results
- On Diverse Driving Weather, Ours(Res101) reaches 55.4 mAP on Day Clear, 42.0 on Night Clear, 39.2 on Dusk Rainy, 24.5 on Night Rainy and 40.6 on Day Foggy, versus 36.9 for C-Gap on the Night Clear domain... with Swin backbone reaching 44.9 average mAP in the table column.
- On Real-to-Art with ResNet-101, Ours reaches 34.8 on Comic, 57.5 on Watercolor and 40.2 on Clipart, outperforming DIV and C-Gap.
- Ablation on Diverse Weather: the one-step prompt baseline gives 32.1 average-style mAP vs 36.2 when adding SDM and 40.6 with all components (Day Foggy column).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace one-step prompting with Chain-of-Thought-Guided Style Evolution (ablation) | Single-DGOD detector · Night Rainy | mAP (%): 14.7 -> 17.9 | Helped |
| Add the Style Disentanglement Module (ablation) | CGSE detector · Night Rainy | mAP (%): 17.9 -> 20.6 | Helped |
| Add Class-Specific Prototype Clustering Module (ablation) | CGSE detector · Night Rainy | mAP (%): 17.9 -> 22.6 | Helped |
| Add both Style Disentanglement Module and Class-Specific Prototype Clustering Module (ablation) | CGSE detector · Night Rainy | mAP (%): 17.9 -> 24.5 | Helped |
| Use a three-tiered chain of thought rather than one-step prompting (ablation) | Style evolution · Diverse Driving Weather Scenarios | mAP (%) | Helped |
| Increase the chain-of-thought hierarchy beyond three tiers (ablation) | Style evolution · Diverse Driving Weather Scenarios | mAP (%) | Hurt |
| Use Ours (Swin) instead of DIV | Single-DGOD detector · Diverse Driving Weather Scenarios — Night Rainy | mAP (%): 24.1 -> 33.7 | Helped |
| Use Ours (Swin) instead of DIV | Single-DGOD detector · Reality-to-Art — Comic | mAP (%): 33.2 -> 36.9 | Helped |
| Use Ours (Swin) instead of DIV | Single-DGOD detector · Reality-to-Art — Watercolor | mAP (%): 57.4 -> 60.7 | Helped |
| Use Ours (Swin) instead of DIV | Single-DGOD detector · Reality-to-Art — Clipart | mAP (%): 38.9 -> 42.5 | Helped |
Surprising findings
- Performance peaks at a three-level chain; adding further levels (level 4 and level 5 sentence templates) reduces performance, which the authors attribute to excessive information.
Limitations
- Gains depend on the quality of the chain-of-thought text prompts; the authors note language-model encoding limits.
- Results use only a single source domain (daytime sunny) and a fixed set of five vocabularies.
Experimental setup
Datasets: S-DGOD driving weather dataset (daytime sunny source) Pascal VOC 2007 and 2012 trainval Clipart1k Watercolor2k Comic2k
Benchmarks: Diverse Driving Weather Scenarios (Day Clear, Night Clear, Dusk Rainy, Night Rainy, Day Foggy) Real-to-Art (Clipart1k, Watercolor2k, Comic2k)
Metrics: mAP@0.5
Baselines: Faster R-CNN SW IBN-Net IterNorm ISW S-DGOD C-Gap PDOC UFR DIV NP
Builds on: AdaIN Faster R-CNN CLIP S-DGOD C-Gap
Closest papers by meaning
- Towards Single-Source Domain Generalized Object Detection via Causal Visual Prompts (NeurIPS 2025 · similarity 0.90)
- Object-Aware Domain Generalization for Object Detection (AAAI 2024 · similarity 0.87)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.87)
- Generalized Diffusion Detector: Mining Robust Features from Diffusion Models for Domain-Generalized Detection (CVPR 2025 · similarity 0.86)
- D3T: Distinctive Dual-Domain Teacher Zigzagging Across RGB-Thermal Gap for Domain-Adaptive Object Detection (CVPR 2024 · similarity 0.85)
- Prompt-Driven Dynamic Object-Centric Learning for Single Domain Generalization (CVPR 2024 · 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.