We summarised the recent dataset for transparent object segmentation (Datasets), and the state-of-the-art segmentation methods (Methods). You can directly access both the papers and their datasets via the super link. 

Transparent object segmentation approaches as shown in the above cover figure can be grouped into two categories: (1) semantic segmentation; (2) instance segmentation. Semantic segmentation approaches that classify pixels with semantic labels are essential for applications in autonomous driving. For example, autonomous robots need to avoid using the unreliable depth information of transparent objects for their self-localisation and avoid collisions with fragile transparent objects during navigation. By extending the scope of semantic segmentation, instance segmentation approaches that partition individual transparent objects are vital for transparent object manipulation tasks, such as pick-and-place.

Given the limited number of instance segmentation approaches for transparent objects, we compare both instance segmentation and semantic segmentation together. Moreover, we only make a detailed comparison of the approaches using deep-learning features, considering the limited pages of this paper.