Learning-Deep-Learning

Light-Head R-CNN: In Defense of Two-Stage Object Detector

April 2019

tl;dr: Faster than two-stage detectors and more accurate than one-stage detectors.

Overall impression

The paper analyzed the computation burden in Faster RCNN and R-FCN, and proposes a more balanced network. The authors fine-tune-fu is amazing.

It is now possible to integrate FPN into R-FCN with the changed architecture of light head RCNN.

The PS RoIPooling is replaced with PS RoIAlign. This RoI Align technique also improved AP by more than 1 point. –> PS RoIAlign is further extended to rotated PS RoIAlign in RoI transformer.

Key ideas

Technical details

Notes