ViCo-LWIR: Vision-Based detection of Indoor Liquid Spills Using Long-Wave Infrared Imaging and a Weakly Supervised Contrastive Framework

Liquid spill detection is a critical concern in safetysensitive indoor environments, such as industrial facilities. When a spill goes unnoticed, it can result in serious consequences, including damage to equipment or products, disruption of operations, and increased safety risks. Traditional cameras that rely on visible light often struggle to detect spills effectively. This is due to the fact that many liquids are clear or semi-transparent, spread in unpredictable patterns, blend into surrounding surfaces, or are difficult to detect under varying lighting conditions. In this paper, we propose a robust liquid detection system using longwave infrared (LWIR) imaging. Despite capable instrumentation for sensing liquid spills, developing a reliable supervised learning system remains a significant challenge. Liquid spills often have unclear boundaries; their thermal behavior can vary depending on the surface, and collecting the detailed pixel-level annotations needed for training detection and segmentation models is both time-consuming and expensive. To address these challenges, we propose a weakly supervised learning framework within a contrastive learning setup. This approach significantly reduces the need for detailed annotations by learning meaningful representations of samples that contain liquid. Our approach remains robust across different surfaces and environmental conditions by explicitly accounting for the uncertainty associated with liquid spills. This objective is achieved by forcing the model to learn grouping semantically similar samples, even when their visual appearances vary through an uncertainty-aware weighting scheme. Samples that do not contain liquid are identified using a hard negative mining strategy. Additionally, training stability and convergence are improved through a curriculum learning scheme that gradually increases the difficulty of the task. Experimental results and comparisons with other weakly supervised methods confirm the robustness and generalization capability of the proposed approach in challenging detection scenarios.