Research map / Detection and segmentation / Open-set and few-shot detection
Triple Feature Disentanglement for One-Stage Adaptive 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 Triple Feature Disentanglement (TFD), a one-stage SSD-based domain adaptive object detector that disentangles domain-invariant and domain-specific features via a multi-level module, cyclic SPP-attention module, and categorical feature module.
Problem
Domain adaptive object detection where simple feature disentanglement leaves residual domain-specific information in domain-invariant representations, and prior disentanglement work is built on two-stage detectors.
Approach
Disentangles features in three phases: a Multi-level Disentanglement Module with a Re-disentanglement Block for DIR, a Cyclic Disentanglement Module with SPP Attention to extract DSR and push it away from DIR, and a Categorical Features Disentanglement Module that aligns per-class features across domains at multiple scales.
Key results
- Achieves 41.2% mAP on Pascal VOC to Clipart1K, up from 40.4% for the prior best (CMT).
- Achieves 55.0% mAP on Pascal VOC to Watercolor2K, up from 53.5% (CMT).
- Achieves 34.4% mAP on Pascal VOC to Comic2K, up from 32.9% (CMT).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add triple feature disentanglement | SSD · Pascal VOC to Clipart1K | mAP: 40.4 -> 41.2 | Helped |
| add triple feature disentanglement | SSD · Pascal VOC to Watercolor2K | mAP: 53.5% to 55.0% | Helped |
| add triple feature disentanglement | SSD · Pascal VOC to Comic2K | mAP: 32.9% to 34.4% | Helped |
| add simple disentanglement (ablation) | SSD · Clipart1K | mAP: 26.7 -> 35.6 | Helped |
| add Re-disentanglement Block (ablation) | SSD with simple disentanglement · Clipart1K | mAP: 35.6 -> 37.9 | Helped |
| add Cyclic Disentanglement Module (ablation) | SSD with RB · Watercolor2K | mAP: 52.2 -> 53.6 | Helped |
| add CFDM part 1 (ablation) | SSD with CDM · Comic2K | mAP: 33.1 -> 33.7 | Helped |
| add CFDM part 2 (ablation) | SSD with CFDM part 1 · Clipart1K | mAP: 39.9 -> 41.2 | Helped |
| use three RBs instead of the full method's RB configuration (ablation) | TFD · Watercolor2K | mAP: 55.0 -> 54.7 | Hurt |
| use two RBs instead of the full method's RB configuration (ablation) | TFD · Clipart1K | mAP: 41.2 -> 41.0 | Hurt |
| set CDM time step to 1 instead of 2 (ablation) | TFD · Watercolor2K | mAP: 55.0 -> 54.2 | Hurt |
| set CDM time step to 3 instead of 2 (ablation) | TFD · Comic2K | mAP: 34.4 -> 34.1 | Hurt |
Surprising findings
- Stacking more Re-disentanglement Blocks degrades performance, which the authors attribute to network capacity limits.
- Time step T=2 is best for the cyclic module; T=3 slightly lowers results.
Limitations
- Improvements over prior methods are modest (about 0.8 to 1.5 mAP points).
- Adding more RB stages hurts performance, limiting scalability of the disentanglement design.
- Evaluated only on Pascal VOC to three cartoon/artistic target domains with SSD.
Experimental setup
Datasets: Pascal VOC Clipart1K Watercolor2K Comic2K
Benchmarks: Clipart1K Watercolor2K Comic2K
Metrics: mAP@0.5
Baselines: Source Only WST BSR BSR+WST SWDA HTCN I3Net DBGL IDF AT LODS CMT
Builds on: SSD VDD orthogonal loss gradient reversal layer SWDA Spatial Pyramid Pooling
Closest papers by meaning
- DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.89)
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (AAAI 2025 · similarity 0.89)
- Non-exemplar Domain Incremental Object Detection via Learning Domain Bias (AAAI 2024 · similarity 0.88)
- DSD-DA: Distillation-based Source Debiasing for Domain Adaptive Object Detection (ICML 2024 · similarity 0.88)
- Differential Alignment for Domain Adaptive Object Detection (AAAI 2025 · similarity 0.88)
- Towards Unsupervised Model Selection for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.87)
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.