SpotNet: Self-Attention Multi-Task Network for Object Detection
SpotNet is an improvement on CenterNet (Keypoint Triplets), which itself is improved based on CornerNet. It’s just like RCNN, Fast RCNN and Faster RCNN, but they are not from the same author / group.
It basically adds two idea:
- Multi-task learning. By asking the model to do both segmentation and bounding box detection, the sahred parameters are more generic and less prone to overfitting.
- Self-Attention. As shown in the image, the result of the segmentation (before binarization) is used as a mask to be applied to the bounding box detection task. This is actually very clever. Q: Where do you get the label for the segmentation task? A: use video and get lower-quality-but-good-enough label by traditional CV method.