Leveraging an Instance Segmentation Method for Detection of Transparent Materials

Abstract

Automatic detection of transparent materials (e.g., glass, plastic, etc.) is essential in many computer vision tasks. For example, a robot could use such a system to navigate around transmissive materials or operate tasks with these materials without causing damage. Nevertheless, it is challenging task as such materials exhibit less texture or background scenes dominate visual perception. Existing methods used either handengineered or leaned features to detect and segment transparent objects. We argue that pixel-wise detection and segmentation of transmissive materials improve detection performance and provide the fine-grained information compared to detecting bounding boxes of objects (i.e., localisation task). In this paper, we leverage a robust and state-of-the-art instance segmentation method namely, Mask R-CNN, in order to detect transparent materials. To be specific, we train the model on a new dataset with an evaluation based on publicly available dataset. Experimental results show that the adopted method significantly enhances the performance of transparent material detection. In particular, the resulting binary masks provides the pixel-level information for an improved understanding and analysis of transparency.

TOSQ: Transparent object segmentation through query based transformer

“…Chen et al [12] introduced TOM-Net, a U-shaped method with consistent feature space dimensions in both encoder and decoder layers. Madessa et al [13] employed Mask R-CNN to detect individual transparent objects. Mei et al [14] proposed GDNet, which utilized a large-field contextual feature integration module and convolutional block attention module (CBAM) for feature fusion.…”