Research map / Detection and segmentation / Open-set and few-shot detection
Modality Translation for Object Detection Adaptation without forgetting prior knowledge
Heitor Rapela Medeiros, Masih Aminbeidokhti, Fidel A Guerrero Pena, David Latortue, Eric Granger, Marco Pedersoli
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| train a small input translator with detection loss instead of image translation (CycleGAN/CUT/HalluciDet) | frozen RGB detector · LLVIP IR test, FCOS | AP: 28.00 -> 57.63 vs HalluciDet | Improved |
| use ModTr instead of HalluciDet with Faster R-CNN on LLVIP | Faster R-CNN · LLVIP IR test | AP: 57.78 -> 57.97 | No clear effect |
| use ModTr instead of HalluciDet on FLIR | FCOS · FLIR IR test | AP: 23.74 -> 35.49 | Improved |
| ModTr with frozen detector instead of full fine-tuning on FLIR | FCOS · FLIR IR test | AP: 27.97 -> 35.49 | Improved |
| ModTr instead of full fine-tuning on LLVIP | Faster R-CNN · LLVIP IR test | AP: 59.62 -> 57.97 | Hurt |
| LoRA on last layer of detector instead of ModTr | LoRA · LLVIP/FLIR | AP: LoRA FCOS LLVIP 47.72 vs ModTr 57.63 | Hurt |
| shrink translator from ResNet34 to MobileNetv2 (ablation) | ModTr_odot · LLVIP, Faster R-CNN | AP: 56.35 -> 56.15 | No clear effect |
| shrink translator from ResNet34 to MobileNetv3s (ablation) | ModTr_odot · FLIR, Faster R-CNN | AP: 37.21 -> 32.06 | Hurt |
| fine-tune a single detector jointly on all modalities (1-Detector) instead of ModTr | detector · COCO val, FCOS | AP: 38.41 -> 00.33 | Hurt |
| use ModTr (frozen detector) rather than N-Detectors for knowledge preservation | FCOS, RetinaNet, Faster R-CNN · Average over LLVIP, FLIR, COCO | AP: FCOS 41.25 -> 43.84 | Improved |
| also fine-tune the detector jointly with ModTr | ModTr_odot · FLIR | AP | Improved |
| add image-quality loss terms to the translator (ablation) | ModTr · LLVIP/FLIR | AP | No clear effect |
Closest papers by meaning
- Visual Modality Prompt for Adapting Vision-Language Object Detectors (ICCV 2025 · similarity 0.87)
- RA-SOD: Reliability-Aware RGB-T Salient Object Detection under Modality Degradation (ECCV 2026 · similarity 0.87)
- MLLM Makes Strong Backbone for Multi-Modal Object Detection (NeurIPS 2026 · similarity 0.86)
- RGB-D Salient Object Detection with Cross-Modality Modulation and Selection (ECCV 2020 · similarity 0.86)
- DUNIT: Detection-Based Unsupervised Image-to-Image Translation (CVPR 2020 · similarity 0.86)
- Keep the Balance: A Parameter-Efficient Symmetrical Framework for RGB+X 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.
