SKU110K: Precise Detection in Densely Packed Scenes

October 2020

tl;dr: object detection in crowded (but not necessarily occluded scene), and proposed one new retail dataset SKU110k.

Overall impression

This paper proposes a novel method to do object detection in densely packed scene with a new NMS method. This method will not necessarily work when there is a lot of occlusion. Fortunately for retail scenes, occlusion does not seem to be a huge problem.

The added IoU branch looks much like that in IoU Net. In both methods, the IoU score (instead of the classification score) is used for bbox NMS. However SKU110k also

This problem may be much better tackled with segmentation or keypoint detection method. In other words, anchor free methods like CenterNet or FCOS may work perfectly in this scenario.

Key ideas

Technical details