Original title: RGB-to-IR image translation for infrared vehicle detection in unseen UAV domains
Thijs A. Eker, Ella P. Fokkinga, Jan Erik van Woerden, Elfi I. S. Hofmeijer +3 · September 2026
Drones often need thermal infrared cameras to spot vehicles, but infrared training pictures are hard to find. Regular color pictures from drones are everywhere. However, converting color pictures to infrared is tough because hidden heat patterns, like warm engines, do not show up in regular photos. Researchers tested whether modern image-generating tools could bridge this gap. They trained translators on paired color and infrared photos, converted regular drone images into fake infrared images, and used them to train vehicle detectors. The fake infrared pictures consistently beat regular color and grayscale images. The top-performing tool was Stable Diffusion 3.5 with ControlNetA technique that guides AI image generators to produce outputs with specific structures or features.. It boosted detection scores on one target group from 50.8 to 60.1, and on another from 25.6 to 38.4. While a gap compared to real infrared data remains, this translation effectively tackles thermal data scarcity.
Imagine trying to learn to identify fruits in a dark room by their shape, but you only have bright, colorful pictures of them. It's hard because you can't see key details like texture or weight. In this paper, researchers create artificial dark-room images from the bright pictures to practice with, so when you actually enter the dark room, you're better prepared. Here, the fake thermal images serve as practice data for detecting vehicles when real thermal images are scarce.
Drone operators for security or wildlife monitoring need to spot vehicles in low-visibility conditions, but real thermal data is hard to get. This method creates more training data without requiring large amounts of real infrared images, making detection systems more effective and accessible.
The synthetic infrared data does not reach the performance of real infrared data, and the study is limited to five aerial datasets, so it may not work equally well in all environments.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
RGB-to-IR image translation for infrared vehicle detection in unseen UAV domains — Thijs A. Eker, Ella P. Fokkinga, Jan Erik van Woerden et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.