README.md

VarifocalNet: An IoU-aware Dense Object Detector

Introduction

VarifocalNet (VFNet) learns to predict the IoU-aware classification score which mixes the object presence confidence and localization accuracy together as the detection score for a bounding box. The learning is supervised by the proposed Varifocal Loss (VFL), based on a new star-shaped bounding box feature representation (the features at nine yellow sampling points). Given the new representation, the object localization accuracy is further improved by refining the initially regressed bounding box. The full paper is available at: https://arxiv.org/abs/2008.13367.

Learning to Predict the IoU-aware Classification Score.

Citing VarifocalNet

@article{zhang2020varifocalnet,
  title={VarifocalNet: An IoU-aware Dense Object Detector},
  author={Zhang, Haoyang and Wang, Ying and Dayoub, Feras and S{\"u}nderhauf, Niko},
  journal={arXiv preprint arXiv:2008.13367},
  year={2020}
}

Results and Models

Backbone	Style	DCN	MS train	Lr schd	Inf time (fps)	box AP (val)	box AP (test-dev)	Config	Download
R-50	pytorch	N	N	1x	-	41.6	41.6	config	model | log
R-50	pytorch	N	Y	2x	-	44.5	44.8	config	model | log
R-50	pytorch	Y	Y	2x	-	47.8	48.0	config	model | log
R-101	pytorch	N	N	1x	-	43.0	43.6	config	model | log
R-101	pytorch	N	Y	2x	-	46.2	46.7	config	model | log
R-101	pytorch	Y	Y	2x	-	49.0	49.2	config	model | log
X-101-32x4d	pytorch	Y	Y	2x	-	49.7	50.0	config	model | log
X-101-64x4d	pytorch	Y	Y	2x	-	50.4	50.8	config	model | log

Notes:

• The MS-train scale range is 1333x[480:960] (range mode) and the inference scale keeps 1333x800.
• DCN means using DCNv2 in both backbone and head.
• Inference time will be updated soon.
• More results and pre-trained models can be found in VarifocalNet-Github