Group Pose: A Simple Baseline for End-to-End Multi-person Pose Estimation

GitHub Link

The GitHub link is https://github.com/michel-liu/grouppose-paddle

Introduce

This repository contains the official PaddlePaddle implementation for the ICCV 2023 paper titled "Group Pose A Simple Baseline for End-to-End Multi-person Pose Estimation." The paper introduces Group Pose, a straightforward transformer-based approach for multi-person pose estimation, treating keypoint prediction as a set of queries. The method simplifies decoder self-attention by using specific group self-attentions instead of interactions between different query types. Experimental results on MS COCO and CrowdPose datasets demonstrate that Group Pose outperforms previous methods without human box supervision, even slightly surpassing ED-Pose, which uses such supervision. The repository provides code, pretrained models, and detailed results for evaluation. The work is released under the Apache 2.0 license. State-of-the-art solutions adopt the DETR-like framework, and mainly develop the complex decoder, e. g., regarding pose estimation as keypoint box detection and combining with human detection in ED-Pose, hierarchically predicting with pose decoder and joint (keypoint) decoder in PETR.

Content

Introduction

In this paper, we study the end-to-end multi-person pose estimation and present a simple yet effective transformer approach, named Group Pose. We simply regard �-keypoint pose estimation as predicting a set of �×� keypoint positions, each from a keypoint query, as well as representing each pose with an instance query for scoring � pose predictions.

Motivated by the intuition that the interaction, among across-instance queries of different types, is not directly helpful, we make a simple modification to decoder self-attention. We replace single self-attention over all the �×(�+1) queries with two subsequent group self-attentions: (i) � within-instance self-attention, with each over � keypoint queries and one instance query, and (ii) (�+1) same-type across-instance self-attention, each over � queries of the same type. The resulting decoder removes the interaction among across-instance type-different queries, easing the optimization and thus improving the performance. Experimental results on MS COCO and CrowdPose show that our approach without human box supervision is superior to previous methods with complex decoders, and even is slightly better than ED-Pose that uses human box supervision.