Pseudo-IoU-for-Anchor-Free-Object-Detection

This is the repo to host the code for Pseudo-IoU in the following paper: Arxiv link

By Jiachen Li, Bowen Cheng, Rogerio Feris, Jinjun Xiong, Thomas S.Huang, Wen-Mei Hwu and Humphrey Shi.

Introduction

Current anchor-free object detectors are quite simple and effective yet lack accurate label assignment methods, which limits their potential in competing with classic anchor-based models that are supported by welldesigned assignment methods based on the Intersectionover-Union (IoU) metric. In this paper, we present Pseudo Intersection-over-Union (Pseudo-IoU): a simple metric that brings more standardized and accurate assignment rule into anchor-free object detection frameworks without any additional computational cost or extra parameters for training and testing, making it possible to further improve anchor-free object detection by utilizing training samples of good quality under effective assignment rules that have been previously applied in anchor-based methods. By incorporating Pseudo-IoU metric into an end-toend single-stage anchor-free object detection framework, we observe consistent improvements in their performanceon general object detection benchmarks such as PASCAL VOC and MSCOCO. Our method (single-model and singlescale) also achieves comparable performance to other recent state-of-the-art anchor-free methods without bells and whistles.

Prerequisites

• Python 3.7
• PyTorch 1.7.0
• CUDA 11.0
• MMdetection v2.11.0

Please following the installation of mmdetection and merges Pseudo-IoU configs and models into mmdetection folder.

Results

More models will be released soon

Backbone	Lr schd	box_mAP	box_mAP_50	box_mAP_75	box_mAP_s	box_mAP_m	box_mAP_l	Config	Download
R-50	1x	38.4	57.4	40.9	23.8	42.5	48.8	config	model
R-101	1x	40.4	59.5	40.9	23.7	44.9	51.4	config	model
R-101-DCN	2x	43.5	62.9	46.6	25.7	47.4	57.6	config	model

Citation

@article{li2021pseudoiou,
  title={Pseudo-IoU: Improving Label Assignment in Anchor-Free Object Detection},
  author={Jiachen Li, Bowen Cheng, Rogerio Feris, Jinjun Xiong, Thomas S.Huang, Wen-Mei Hwu and Humphrey Shi},
  journal={arXiv preprint arXiv:2104.14082},
  year={2021}
}