Embedded Files
Figure 1. Sample objects from the TRANS-AFF Dataset. Top row: RGB images. Middle row: Depth images. Bottom row: Pixel-wise affordances label. Red, blue and green colour represent "contain'', "wrap-grasp'' and "support'' respectively.
TABLE I: Description of the three affordance labels
Data Collection Tool
Data Collection Tool
We use the data collection tool from ClearGrasp paper. To provide accurate ground-truth depth maps for the transparent objects, we collected a “twin” RGB-D pair for every scene that contains transparent objects. In the “twin” setup, the transparent objects in the original image are replaced with an identical spray-painted instance that can reflect light evenly and provide accurate depth information
Data Annotation Tool
Data Annotation Tool
We use an instance segmentation annotation tool to label the affordance map of transparent objects. The software can be found in the following GitHub link. First, we need to annotate instance mask for every object, then we need to annotate the "contain", "support" and "wrap-grasp" affordance map for every instance, which will output a coco-format file. Finally, we write a program to change the coco-format file to a set of affordance maps for the images used in our dataset.
Page updated
Google Sites
Report abuse