Instance segmentation is a computer vision technique utilized to detect and separate individual objects in an image. Within the realm of warehouse automation, instance segmentation is used to identify and define distinct objects that are stored in containers. The outcomes of instance segmentation can be used to provide information to subsequent robotic processes, such as the identification of objects and generation of grasping strategies. It is an essential factor in achieving high-efficiency robot manipulation as it facilitates more precise and accurate object detection and handling.
We have divided the Object Segmentation dataset into three distinct subsets to showcase the challenge of transferring learned knowledge across tasks. The subsets are as follows:
- Mix-Object-Tote (14G): This subset consists of close-up images of mixed objects that are stored in either yellow or blue totes. Mix-Object-Tote comprises a total of 44,253 images of size 2448 by 2048 pixels and 467,225 annotations, with an average of 10.5 instances per tote.
- Zoomed-Out-Tote-Transfer-Set (1.5G): This subset includes mixed objects placed in a yellow tote that were captured with sensors positioned further away from the tote, under different lighting conditions. The Zoomed-Out-Tote-Transfer-Set contains 5,837 images of size 2046 by 2046 pixels and 43,401 annotations, with an average of 7.5 objects per tote.
- Same-Object-Transfer-Set (3G): This subset consists of multiple same objects placed in close proximity within various storage units. The Same-Object-Transfer-Set comprises 3,323 images of size 2048 by 1500 pixels and 12,664 annotations, with an average of 3.8 objects per scene.
These labels are provided per-image in the format that is compatible with LabelMe, as well as a train, validation, and test split format that is compatible with MS-COCO format.