Original title: ProG-Net: Prompted Guidance Network for Visible–Thermal Tiny-Object Detection
Yan Zhang, Wang Qiang, Hui Li · September 2026
Cameras that combine regular visible light with thermal heat vision help self-driving machines see in all weather. However, spotting tiny objects is very hard. These targets have almost no visible texture and vanish when computers shrink images to analyze them. Different camera types also record very different signals. To solve this, researchers built ProG-Net. It starts with a frozen, pre-trained visual foundation model to borrow strong visual knowledge. Next, a special helper branch predicts specific target points. These point prompts act as guides, steering the model to focus on tiny objects instead of background clutter. Finally, custom fusion modules combine and align both vision streams to fix data imbalances. Testing showed ProG-Net beat top competing systems across multiple scores.
You are searching for a lost earring in a large, cluttered room. Normally, you scan the whole room once and miss tiny things like earrings because your eyes skip over them. Now imagine you have a helper who points exactly to the spot where the earring is — you only need to look there. In this paper, the detection system works like that helper: it uses a point prompt to direct the frozen AI model's attention to the tiny object's location, while also combining visible and thermal images at that spot to confirm what it is.
Autonomous cars and surveillance drones need to spot small objects like distant pedestrians or drones in any weather. Visible-thermal systems struggle when those objects are tiny — they vanish in digital processing. ProG-Net solves this by using a frozen AI model that targets tiny regions precisely, making detection reliable where it previously failed.
The abstract does not discuss limitations explicitly. It evaluates on two specific benchmarks (RGBT-Tiny and LLVIP), so generalizability to other datasets or real-world conditions with different sensor setups is not established. The reliance on a frozen VFM backbone may also prevent adaptation to novel target types.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
ProG-Net: Prompted Guidance Network for Visible–Thermal Tiny-Object Detection — Yan Zhang, Wang Qiang, Hui Li — Remote Sensing — doi:10.3390/rs18183239
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.