Cascade R-CNN: Delving into High Quality Object Detection
The basic idea of this paper is to train multiple detection heads with multiple IoU thresholds. The output of the previous detector is fed to the next as a resampling mechanism.
This method gives better results, especially for high IoUs.
This approach is somewhat expensive as it adds 100M parameters and is slower by 0.03 seconds on a FPN during inference. (0.115s vs 0.14)
Code is available here : https://github.com/zhaoweicai/cascade-rcnn