UniHead

Official code for "Unifying Visual Perception by Dispersible Points Learning". The implementation is based on United-Percepion.

Introduction

UniHead is a plug-in perception head which can be used in different detection frameworks (two-stage or one-stage pipelines), and different tasks (image classification, object detection, instance segmentation and pose estimation).

Guide to Our Code

Currently, configs can be found in configs/unihead.

Experiments on MS-COCO 2017

Our original implementation is based on the unreleased internal detection framework so there may be a small performance gap.

Different Detection Pipelines

Pipeline	mAP	Config	Model
two-stage	42.0	config	google
cascade	42.8	config	google

Different Tasks

Task	mAP	Config	Model
detection	42.0	config	google
instance segmentation	30.3	config	google
pose estimation	57.6	config	google

More results and models will soon be released.

LICENSE

This project is released under the MIT license. Please see the LICENSE file for more information.

Citation

@article{liang2022unifying,
  author  = {Jianming Liang, Guanglu Song, Biao Leng, Yu Liu},
  journal = {arXiv:2208.08630},
  title   = {Unifying Visual Perception by Dispersible Points Learning},
  year    = {2022},
}