YuChuang1205/PAL

What it solves

This project addresses the challenge of Infrared Small Target Detection (SIRST) when high-quality, full-pixel annotations are unavailable. It specifically focuses on training effective detection models using only single-point supervision (coarse or centroid points), which is much easier to obtain than full masks but typically leads to lower performance.

How it works

PAL implements a Progressive Active Learning framework based on the philosophy of "easy to hard" learning. Instead of treating all training samples equally, the system progressively identifies and prioritizes the learning of harder samples based on the current model's capabilities. This allows the network to continuously enhance its performance by actively recognizing and adapting to more complex targets over time.

Who it’s for

This framework is designed for researchers and engineers working on infrared target detection, computer vision for surveillance or defense, and those looking to implement SIRST models with limited annotation budgets.

Highlights

• Broad Model Support: Integrates over ten existing SIRST detection networks (e.g., MSDA-Net, ISNet, DNANet) into a single framework.
• Flexible Supervision: Supports three types of label forms: full supervision, coarse single-point supervision, and centroid single-point supervision.
• Dataset Compatibility: Works with multiple standard datasets including SIRST, NUDT-SIRST, IRSTD-1k, and SIRST3.
• Easy Configuration: Allows switching between different networks, datasets, and label forms by changing a single parameter in the training script.