Research map / Detection and segmentation / Open-set and few-shot detection
Caltech Aerial RGB-Thermal Dataset in the Wild
Connor Lee, Matthew Anderson, Nikhil Ranganathan, Xingxing Zuo, Kevin T Do, Georgia Gkioxari, Soon-Jo Chung
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 |
|---|---|---|---|
| use larger model (ConvNext-B) instead of fastest FastSCNN | thermal semantic segmentation · general split | mIoU: +0.027 mIoU at 130x FLOPs | No clear effect |
| use frozen large RGB foundation models (DINOv2, ConvNext-B CLIP) instead of ImageNet-pretrained lightweight nets | thermal segmentation · general split | mIoU | No clear effect |
| use thermal-specific FTNet edge loss instead of cross entropy baselines | thermal segmentation · general split | mIoU | Hurt |
| train on one region/terrain and test on others (ablation) | EfficientViT-B0 · geographic split | avg mIoU: 0.314 (train CA) vs 0.586 (train All) | Hurt |
| add ground-captured images to aerial training data (ablation) | EfficientViT-B0 · terrain split | avg mIoU: 0.518 -> 0.532 | Improved |
| train on daytime only and test at sunrise or night (ablation) | EfficientViT-B0 · temporal split | mIoU: 0.777 (day) -> 0.242 (sunrise), 0.193 (night) | Hurt |
| add RGB to thermal input in RGB-T fusion models | semantic segmentation · RGB-T split | mIoU: 0.661 (thermal only) -> 0.740 (CMNeXt) | Improved |
| use paired GANs instead of unpaired GANs or diffusion for RGB-to-thermal translation | image translation · RGB to thermal | PSNR/SSIM/mIoU: thermal mIoU 0.189 (UNIT) -> 0.388 (VQGAN) | Improved |
| use thermal imagery instead of RGB for SAM and vision-language zero-shot segmentation | foundation models · zero-shot | mIoU / AP: ODISE mIoU 0.504 (color) -> 0.232 (thermal); SAM AP 0.538 -> 0.018 | Hurt |
| pretrain with MoCoV2 on urban thermal data instead of ImageNet weights (ablation) | thermal segmentation networks · Caltech Aerial test set | mIoU: EfficientViT 0.725 -> 0.714 | No clear effect |
| use ImageNet pretraining instead of training from scratch (ablation) | thermal segmentation networks · Caltech Aerial test set | mIoU: EfficientViT 0.687 -> 0.725; FastSCNN 0.640 -> 0.690 | Improved |
| use thermal instead of RGB input for VINS-Fusion | VINS-Fusion · aerial sequences | ATE: North Field N1 2.883 (RGB) vs 9.454 (thermal); Castaic Lake C2 --- vs 1.513 | Mixed |
Closest papers by meaning
- ThermalGen: Style-Disentangled Flow-Based Generative Models for RGB-to-Thermal Image Translation (NeurIPS 2025 · similarity 0.89)
- Unleashing Multispectral Video's Potential in Semantic Segmentation: A Semi-supervised Viewpoint and New UAV-View Benchmark (NeurIPS 2024 · similarity 0.84)
- USCILab3D: A Large-scale, Long-term, Semantically Annotated Outdoor Dataset (NeurIPS 2024 · similarity 0.84)
- High Temporal Consistency through Semantic Similarity Propagation in Semi-Supervised Video Semantic Segmentation for Autonomous Flight (CVPR 2025 · similarity 0.83)
- MCD: Diverse Large-Scale Multi-Campus Dataset for Robot Perception (CVPR 2024 · similarity 0.83)
- ThermalGaussian: Thermal 3D Gaussian Splatting (ICLR 2025 · similarity 0.83)
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.
