PreciseRoIPooling

Why we included this project

PreciseRoIPooling implements the PrRoI pooling operator from the ECCV 2018 paper on localization confidence, and it addresses a real annoyance for anyone training two-stage detectors in PyTorch. Instead of quantizing RoIs into discrete bins like Fast R-CNN's RoI pooling, or sampling a fixed set of points like RoI Align, it integrates over the region with bilinear interpolation, which gives a continuous gradient with respect to the box coordinates. That means the loss can refine box positions rather than treating them as fixed inputs. The implementation is small and self-contained, PyTorch-only, and the paper spells out the gradient math if you need it. One practical gotcha: install by cloning the repo, since the zip download breaks the symbolic links in the source tree.