Influence of Water Droplet Contamination for Transparency Segmentation

TL;DR

This work studies segmentation of transparent objects under water droplet contamination and introduces a real-world dataset of 489 images with three contamination grades. It evaluates a transformer-based transparency segmentation model (Trans4Trans) and a foundation model (SAM) with pretraining on Trans10K and subsequent transfer-training on the new dataset. The results show contaminated transparent objects are easier to segment and that contamination level is detectable, enabling cleaning alerts and improved anomaly resilience, with IoU gains of approximately 7.4 and 6.4 percentage points across contamination transitions. IoU improvements are achieved through a transfer-learning pipeline that pretrains on Trans10K with a PVT-Medium backbone and adapts to the contamination data, providing a publicly usable benchmark for robust transparency segmentation in contamination-prone industrial settings. Overall, the work offers a practical path toward resilient vision systems in environments where transparency and contamination interact, along with a dataset to benchmark future methods.

Abstract

Computer vision techniques are on the rise for industrial applications, like process supervision and autonomous agents, e.g., in the healthcare domain and dangerous environments. While the general usability of these techniques is high, there are still challenging real-world use-cases. Especially transparent structures, which can appear in the form of glass doors, protective casings or everyday objects like glasses, pose a challenge for computer vision methods. This paper evaluates the combination of transparent objects in conjunction with (naturally occurring) contamination through environmental effects like hazing. We introduce a novel publicly available dataset containing 489 images incorporating three grades of water droplet contamination on transparent structures and examine the resulting influence on transparency handling. Our findings show, that contaminated transparent objects are easier to segment and that we are able to distinguish between different severity levels of contamination with a current state-of-the art machine-learning model. This in turn opens up the possibility to enhance computer vision systems regarding resilience against, e.g., datashifts through contaminated protection casings or implement an automated cleaning alert.

Figures (3)

• Figure 1: Overview of the different scenes captured for the dataset.
• Figure 2: Example scenes where the application of contamination could influence the segmentation quality in a meaningful way. The level of contamination is encoded through the colors of the ground truth annotations, with white denoting no contamination, while red denotes strong contamination. For Scene 1, presence of contamination did improve the segmentation, whereas for Scene 2, our simulation degraded the segmentation quality. This happens most likely through the fence structure, which is very similar to a structure that can encapsulate a glass pane.
• Figure 3: Examples for scenes in which the model was able to properly segment the type of contamination applied to the surface. A white segmentation denotes the prediction of no contamination, yellow the prediction of little contamination and red the detection of strong contamination. In Scene 1, the model failed to fully distinguish between little and strong contamination in the marked areas.