VAP-SAM: Prompt-Free SAM for Insubstantial Gas Segmentation

Insubstantial gas targets play an important role in industrial production and medical diagnostics. However, the pixel-level segmentation of gas leakage locations remains a largely unexplored task. Unlike rigid objects, insubstantial gas targets lack fixed shapes, exhibit highly variable boundaries, and often present blurred appearances due to inconsistent concentrations. These characteristics make Insubstantial Gas Segmentation an inherently challenging task. Moreover, there is currently no publicly available dataset dedicated to this research area. To address this gap, we introduce Gasseg-ZSL, the first open-source dataset for insubstantial gas segmentation, and propose VAP-SAM for pixel-level gas detection. To enhance edge-awareness of gas regions, we design Vision-Autonomous Perception (VAP) module built upon HQ-SAM. VAP generates multiple perception masks via object queries, followed by a confidence query to select the mask with the highest score. Furthermore, VAP is pretrained with object detection to activate its perception capability during bounding-box regression. Finally, we employ Prior Knowledge Distillation, where a prompt-based teacher model provides soft labels to guide the student model VAP-SAM, enabling it to capture implicit prompt priors for gas segmentation. Despite introducing only 0.91M additional parameters, VAP-SAM achieves significant performance gains in a prompt-free setting, boosting BIoU, IoU, and Dice by 13.55%, 10.05%, and 7.2%, respectively. Remarkably, it also outperforms HQ-SAM trained with explicit prompt priors. Gasseg-ZSL dataset is available at https://github.com/zhushiL/Gasseg-ZSL.