Research map / Detection and segmentation / Open-set and few-shot detection
D3T: Distinctive Dual-Domain Teacher Zigzagging Across RGB-Thermal Gap for Domain-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 D3T, a mean-teacher UDA framework with separate RGB and thermal teachers and a zigzag schedule that alternates domain-specific training to adapt object detectors from RGB to thermal images.
Problem
Unsupervised domain adaptation from visible (RGB) to thermal images for object detection suffers from a much larger domain gap than RGB-to-RGB adaptation, and standard single-teacher methods fail to bridge it.
Approach
Use two distinct EMA teachers (RGB and thermal) that are each updated only when training on their own domain, while the student trains on one domain at a time and learns from both teachers' pseudo-labels; a zigzag schedule shifts training frequency from RGB to thermal over time, and pseudo-label weight lambda is ramped from 0 to 1.
Key results
- On FLIR RGB-to-thermal adaptation, D3T reaches 69.30 mAP versus 65.81 for HT and 34.68 for source-only.
- On KAIST RGB-to-thermal adaptation, D3T reaches 48.96 person AP versus 43.45 for HT.
- Zigzag scheduling gives 69.30 mAP versus 68.28 for the best fixed-iteration setting (Table 5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use D3T instead of HT | FCOS with VGG-16 · RGB→Thermal FLIR | mAP: 65.81 -> 69.30 | Helped |
| use D3T instead of HT | FCOS with ResNet-50 · RGB→Thermal KAIST | Person AP: 43.45 -> 48.96 | Helped |
| add Dual-Teachers (ablation) | D3T on FLIR · RGB→Thermal FLIR | mAP: 65.81 -> 66.93 | Helped |
| add Zigzag-Learn to Dual-Teachers (ablation) | D3T on FLIR · RGB→Thermal FLIR | mAP: 66.93 -> 68.46 | Helped |
| add Incor-Know to Dual-Teachers and Zigzag-Learn (ablation) | D3T on FLIR · RGB→Thermal FLIR | mAP: 68.46 -> 69.30 | Helped |
| replace fixed 50-iteration zigzag with dynamic zigzag (ablation) | D3T on FLIR · RGB→Thermal FLIR | mAP: 65.57 -> 69.30 | Helped |
| replace fixed 100-iteration zigzag with dynamic zigzag (ablation) | D3T on FLIR · RGB→Thermal FLIR | mAP: 68.28 -> 69.30 | Helped |
| replace fixed 1,000-iteration zigzag with dynamic zigzag (ablation) | D3T on FLIR · RGB→Thermal FLIR | mAP: 65.36 -> 69.30 | Helped |
| use fixed pseudo-label weight λ=1 instead of no pseudo-label loss (ablation) | D3T RGB-domain training · RGB→Thermal FLIR | mAP: 68.46 -> 55.12 | Hurt |
| use fixed pseudo-label weight λ=0.1 instead of no pseudo-label loss (ablation) | D3T RGB-domain training · RGB→Thermal FLIR | mAP: 68.46 -> 68.57 | Helped |
| dynamically increase pseudo-label weight from 0 to 1 instead of using no pseudo-label loss (ablation) | D3T RGB-domain training · RGB→Thermal FLIR | mAP: 68.46 -> 69.30 | Helped |
Surprising findings
- Training the student with pseudo-labels at full weight (lambda=1) drops FLIR mAP to 55.12, far below using no pseudo-labels (68.46).
Limitations
- Evaluated only on two thermal datasets (FLIR, KAIST) with a custom disjoint-image protocol, and only with FCOS.
- Relies on paired-free but hand-tuned schedule hyperparameters (Z_thr, Z_rgb, beta, lambda).
Experimental setup
Datasets: FLIR KAIST
Benchmarks: FLIR KAIST
Metrics: mAP
Baselines: Source only DANN SWDA EPM HT
Builds on: Mean Teacher HT (H-Teacher) FCOS
Closest papers by meaning
- CAT: Exploiting Inter-Class Dynamics for Domain Adaptive Object Detection (CVPR 2024 · similarity 0.88)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.88)
- Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection (NeurIPS 2024 · similarity 0.88)
- Triple Feature Disentanglement for One-Stage Adaptive Object Detection (AAAI 2024 · similarity 0.87)
- Mean Teacher DETR with Masked Feature Alignment: A Robust Domain Adaptive Detection Transformer Framework (AAAI 2024 · similarity 0.87)
- DA-Ada: Learning Domain-Aware Adapter 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.