An Implementation of Faster RCNN with Study for Region Sampling

TL;DR

This paper presents a TensorFlow implementation of Faster R-CNN, analyzing region sampling methods and demonstrating that biased sampling towards small regions can match NMS-based sampling performance.

Contribution

It provides a TensorFlow baseline implementation of Faster R-CNN and investigates the impact of different region sampling strategies on detection performance.

Findings

Biased small region sampling improves detection accuracy.

The implementation is publicly available for further research.

Small region sampling can match NMS-based sampling when converged.

Abstract

We adapted the join-training scheme of Faster RCNN framework from Caffe to TensorFlow as a baseline implementation for object detection. Our code is made publicly available. This report documents the simplifications made to the original pipeline, with justifications from ablation analysis on both PASCAL VOC 2007 and COCO 2014. We further investigated the role of non-maximal suppression (NMS) in selecting regions-of-interest (RoIs) for region classification, and found that a biased sampling toward small regions helps performance and can achieve on-par mAP to NMS-based sampling when converged sufficiently.