基于改进HRNet的高精度鱼类姿态估计方法

doi:10.12133/j.smartag.SA202502001

Smart Agriculture ›› 2025, Vol. 7 ›› Issue (3): 160-172.doi: 10.12133/j.smartag.SA202502001

• 信息处理与决策 • 上一篇

基于改进HRNet的高精度鱼类姿态估计方法

彭秋珺¹^,²^,³^,⁴^,⁵, 李蔚然¹^,²^,³^,⁴^,⁵, 刘业强¹^,²^,³^,⁴^,⁵, 李振波¹^,²^,³^,⁴^,⁵()

^1. 中国农业大学国家数字渔业创新中心，北京 100083，中国
^2. 中国农业大学信息与电气工程学院，北京 100083，中国
^3. 农业农村部智慧养殖技术重点实验室，北京 100083，中国
^4. 农业农村部信息化标准化重点实验室，北京 100083，中国
^5. 北京市农业物联网工程技术研究中心，北京 100083，中国

收稿日期:2025-01-25 出版日期:2025-05-30
基金项目:
国家重点研发计划项目(2020YFD0900204); 新一代人工智能国家科技重大专项(2021ZD0113805); 北京市智慧农业创新团队项目(BAIC10-2024)
作者简介:
彭秋珺，硕士研究生，研究方向为计算机视觉。E-mail：pengqiujun@cau.edu.cn
通信作者:
李振波，博士，教授，研究方向为计算机视觉。E-mail：lizb@cau.edu.cn

High-Precision Fish Pose Estimation Method Based on Improved HRNet

PENG Qiujun¹^,²^,³^,⁴^,⁵, LI Weiran¹^,²^,³^,⁴^,⁵, LIU Yeqiang¹^,²^,³^,⁴^,⁵, LI Zhenbo¹^,²^,³^,⁴^,⁵()

^1. National Innovation Center for Digital Fishery, China Agricultural University, Beijing 100083, China
^2. College of Information and Electrical Engineering, China Agricultural University, Beijing 100083, China
^3. Key Laboratory of Smart Farming for Aquatic Animal and Livestock, Ministry of Agriculture and Rural Affairs, Beijing 100083, China
^4. Key Laboratory of Agricultural Informatization Standardization, Ministry of Agriculture and Rural Affairs, Beijing, 100083, China
^5. Beijing Engineering and Technology Research Center for Internet of Things in Agriculture, Beijing 100083, China

Received:2025-01-25 Online:2025-05-30
Foundation items:National Key Research and Development Program of China(2020YFD0900204); National Science and Technology Major Project(2021ZD0113805); Beijing Smart Agriculture Innovation Consortium Project(BAIC10-2024)
About author:
PENG Qiujun, E-mail: pengqiujun@cau.edu.cn
Corresponding author:
LI Zhenbo, E-mail: lizb@cau.edu.cn

摘要/Abstract

摘要：

【目的/意义】 鱼类姿态估计是获取鱼类生理信息的重要手段，对于水产养殖中的健康监测具有重要意义。鱼类在受到损伤时通常表现出异常行为，并伴随身体部位位置的变化。此外，鱼类游动过程中的姿态和方向具有不确定性，且游速较快，这对鱼类姿态估计研究提出了挑战。因此，本研究提出了一种名为HPFPE（High-Precision Fish Pose Estimation）的鱼类姿态估计模型，旨在精准捕捉鱼类姿态并准确识别其关键点。 【方法】 该模型一方面将注意力机制模块（Convolutional Block Attention Module， CBAM）集成到HRNet（High-Resolution Net）框架中，在不增加计算复杂度的前提下提高模型预测精度。另一方面，模型通过融合空洞卷积来扩大感受野，提取更广泛的空间上下文信息，从而进一步提升姿态估计的准确性。 【结果和讨论】 在斑石鲷数据上，与HRNet方法相比，HPFPE在不同骨干网络和输入尺寸下的平均准确率分别提高了0.62、1.35、1.76和1.28个百分点，平均召回率分别提升了0.85、1.50、1.40和1.00个百分点。此外，HPFPE在性能上也优于DeepPose、CPM（Convolutional Pose Machine）、SCNet（Self-Calibrated Convolutions Net）和Lite-HRNet。在观赏鱼数据上，HPFPE的平均精度和平均召回率分别达到52.96%和59.50%，显著优于其他对比方法。 【结论】 本研究提出的HPFPE模型能够有效估计鱼类姿态，为鱼类行为识别等应用提供重要参考。

关键词: 水产养殖, 计算机视觉, 鱼类姿态估计, 关键点, 注意力机制

Abstract:

[Objective] Fish pose estimation (FPE) provides fish physiological information, facilitating health monitoring in aquaculture. It aids decision-making in areas such as fish behavior recognition. When fish are injured or deficient, they often display abnormal behaviors and noticeable changes in the positioning of their body parts. Moreover, the unpredictable posture and orientation of fish during swimming, combined with the rapid swimming speed of fish, restrict the current scope of research in FPE. In this research, a FPE model named HPFPE is presented to capture the swimming posture of fish and accurately detect their key points. [Methods] On the one hand, this model incorporated the CBAM module into the HRNet framework. The attention module enhanced accuracy without adding computational complexity, while effectively capturing a broader range of contextual information. On the other hand, the model incorporated dilated convolution to increase the receptive field, allowing it to capture more spatial context. [Results and Discussions] Experiments showed that compared with the baseline method, the average precision (AP) of HPFPE based on different backbones and input sizes on the oplegnathus punctatus datasets had increased by 0.62, 1.35, 1.76, and 1.28 percent point, respectively, while the average recall (AR) had also increased by 0.85, 1.50, 1.40, and 1.00, respectively. Additionally, HPFPE outperformed other mainstream methods, including DeepPose, CPM, SCNet, and Lite-HRNet. Furthermore, when compared to other methods using the ornamental fish data, HPFPE achieved the highest AP and AR values of 52.96%, and 59.50%, respectively. [Conclusions] The proposed HPFPE can accurately estimate fish posture and assess their swimming patterns, serving as a valuable reference for applications such as fish behavior recognition.

Key words: aquaculture, computer vision, fish pose estimation, key point, attention mechanism

中图分类号:

S951.2

彭秋珺, 李蔚然, 刘业强, 李振波. 基于改进HRNet的高精度鱼类姿态估计方法[J]. 智慧农业(中英文), 2025, 7(3): 160-172.

PENG Qiujun, LI Weiran, LIU Yeqiang, LI Zhenbo. High-Precision Fish Pose Estimation Method Based on Improved HRNet[J]. Smart Agriculture, 2025, 7(3): 160-172.

图/表 19

参考文献 26

[1]	PENG Q, LI W, LI Z. Review of fish behavior recognition methods based on artificial intelligence[J]. Transactions of the Chinese society for agricultural machinery, 2023, 54(S1): 283-295.
[2]	ZHENG C, WU W H, CHEN C, et al. Deep learning-based human pose estimation: A survey[J]. ACM computing surveys, 2024, 56(1): 1-37.
[3]	SUN K, XIAO B, LIU D, et al. Deep high-resolution representation learning for human pose estimation[C]// 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2019: 5686-5696.
[4]	LIU J J, HOU Q B, CHENG M M, et al. Improving convolutional networks with self-calibrated convolutions[C]// 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2020: 10093-10102.
[5]	YU C Q, XIAO B, GAO C X, et al. Lite-HRNet: A lightweight high-resolution network[C]// 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2021: 10435-10445.
[6]	YAN S J, XIONG Y J, LIN D H. Spatial temporal graph convolutional networks for skeleton-based action recognition[C]// Proceedings of the AAAI conference on artificial intelligence. Palo Alto, California USA: AAAI Press. 2018, 32(1): 7444-7452.
[7]	MARKOVITZ A, SHARIR G, FRIEDMAN I, et al. Graph embedded pose clustering for anomaly detection[C]// 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2020: 10536-10544.
[8]	DAS S, SHARMA S, DAI R, et al. VPN: Learning video-pose embedding for activities of daily living[C]// Computer Vision – ECCV 2020. Cham, Germany: Springer International Publishing, 2020: 72-90.
[9]	YANG X T, ZHANG S, LIU J T, et al. Deep learning for smart fish farming: Applications, opportunities and challenges[J]. Reviews in aquaculture, 2021, 13(1): 66-90.
[10]	YUE J, YANG H H, JIA S X, et al. A multi-scale features-based method to detect Oplegnathus [J]. Information processing in agriculture, 2021, 8(3): 437-445.
[11]	WANG H R, CHEN Y Y, CHAI Y Q, et al. Image segmentation method combined with VoVNetv2 and shuffle attention mechanism for fish feeding in aquaculture[J]. Smart agriculture, 2023, 5(4): 137-149.
[12]	LI W R, LI F, LI Z B. CMFTNet: Multiple fish tracking based on counterpoised JointNet[J]. Computers and electronics in agriculture, 2022, 198: ID 107018.
[13]	ZHAO Y Y, LI W R, LI Y M, et al. LFCNet: A lightweight fish counting model based on density map regression[J]. Computers and electronics in agriculture, 2022, 203: ID 107496.
[14]	BANAN A, NASIRI A, TAHERI-GARAVAND A. Deep learning-based appearance features extraction for automated carp species identification[J]. Aquacultural engineering, 2020, 89: ID 102053.
[15]	DU Z Z, XU X B, BAI Z Z, et al. Feature fusion strategy and improved GhostNet for accurate recognition of fish feeding behavior[J]. Computers and electronics in agriculture, 2023, 214: ID 108310.
[16]	LI B B, YUE J, JIA S X, et al. Recognition of abnormal body surface characteristics of Oplegnathus punctatus [J]. Information processing in agriculture, 2022, 9(4): 575-585.
[17]	HU W C, YANG X T, MA P C, et al. DCA-MVIT: Fused DSGated convolution and CA attention for fish feeding behavior recognition in recirculating aquaculture systems[J]. Aquaculture, 2025, 598: ID 742008.
[18]	MATHIS A, MAMIDANNA P, CURY K M, et al. DeepLabCut: Markerless pose estimation of user-defined body parts with deep learning[J]. Nature neuroscience, 2018, 21(9): 1281-1289.
[19]	WU X, HUANG J P, WANG L M. Pose estimation-based visual perception system for analyzing fish swimming[J]. IEEE sensors journal, 2024, 24(8): 13293-13303.
[20]	WANG L, ZOU H Q, CHEN S H. Fish identification and tracking based on pose estimation[C]// 2024 IEEE 19th Conference on Industrial Electronics and Applications (ICIEA). Piscataway, New Jersey, USA: IEEE, 2024.
[21]	CUI H P, QIN C X, MA Z Y. Fish key feature point detection and sign identification based on deep learning[J]. Journal of Chinese agricultural mechanization, 2024, 45(6): 201-207.
[22]	YU F, KOLTUN V. Multi-scale context aggregation by dilated convolutions[EB/OL]. arXiv: 1511.07122, 2015.
[23]	WOO S, PARK J, LEE J Y, et al. CBAM: Convolutional Block attention module[C]// Computer Vision – ECCV 2018. Cham, Germany: Springer International Publishing, 2018: 3-19.
[24]	KAN Z H, CHEN S S, ZHANG C, et al. Self-correctable and adaptable inference for generalizable human pose estimation[C]// 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2023: 5537-5546.
[25]	HUANG L Z, LI Y L, TIAN H B, et al. Semi-supervised 2D human pose estimation driven by position inconsistency pseudo label correction module[C]// 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2023: 693-703.
[26]	WANG H X, ZHOU L, CHEN Y Y, et al. Regularizing vector embedding inBottom-up human pose estimation[C]// Computer Vision – ECCV 2022. Cham, Germany: Springer, 2022: 107-122.

Number	Name	Location
1	belly	Anterior end of the pelvic fin
2	eye_left	Left eye
3	eye_right	Right eye
4	back	Anterior end of dorsal fin
5	tail_top	Anterior end of tail
6	tail_up	Top of tail
7	tail_low	Bottom of tail

Dataset	Training set images	Test set images	Image total	Fish total
Oplegnathus punctatus images before augmentation	400	100	500	800
Oplegnathus punctatus images after augmentation	800	100	900	1 800
Ornamental fish images	240	60	300	3 000
Final dataset total	1 040	160	1 200	4 800

Layer	Multi-scale feature map	Number of branches	Number of blocks in each branch	Number of exchange units
Stage 1	1/4	1	4	0
Stage 2	1/4， 1/8	2	4， 4	1
Stage 3	1/4， 1/8， 1/16	3	4， 4， 4	4
Stage 4	1/4， 1/8， 1/16， 1/32	4	4， 4， 4， 4	3

Configuration	Parameter
CPU	Intel（R） Xeon（R） Silver 4210R CPU
GPU	NVIDIA Tesla V100
Operating system	Ubuntu 18.04.6 LTS
Deep learning framework	PyTorch 1.8.1
Programming language	Python 3.8

Number of stage	HRNet-32	HRNet-48
Stage 2	64	96
Stage 3	128	192
Stage 4	256	384

基于改进HRNet的高精度鱼类姿态估计方法

High-Precision Fish Pose Estimation Method Based on Improved HRNet

在线阅读

知网下载

本地下载

可视化

摘要/Abstract

引用本文

使用本文

图/表 19

参考文献 26

相关文章 15

编辑推荐

Metrics

本文评价

Method	Backbone	Input size	AP/%	AP₅₀/%	AP₇₅/%	AR/%	AR₅₀/%	AR₇₅/%
HRNet	HRNet-W32	256×192	71.50	96.72	76.69	75.45	97.00	79.00
	HRNet-W32	384×288	72.70	96.95	78.77	76.35	97.50	81.00
	HRNet-W48	256×192	71.15	94.53	77.39	75.25	95.50	80.50
	HRNet-W48	384×288	72.84	96.66	80.74	76.60	97.00	82.50
HPFPE (Ours)	HRNet-W32	256×192	72.12	97.98	74.87	76.30	98.50	78.50
	HRNet-W32	384×288	74.05	99.01	75.82	77.85	99.50	80.00
	HRNet-W48	256×192	72.91	96.90	78.13	76.65	97.00	80.50
	HRNet-W48	384×288	74.12	97.89	81.99	77.60	98.00	84.00

Method	Backbone	AP/%	AP₅₀/%	AP₇₅/%	AR/%	AR₅₀/%	AR₇₅/%
DeepPose	ResNet-50	42.24	83.66	36.89	57.30	90.00	58.50
	ResNet-101	41.46	85.08	35.71	57.05	91.00	58.00
	ResNet-152	43.04	85.46	37.07	58.40	91.00	59.00
CPM	CPM	62.22	93.53	67.25	66.45	95.00	71.00
SCNet	SCNet-50	70.15	96.86	74.54	73.95	97.00	78.00
SCNet	SCNet-101	69.45	95.78	74.36	74.10	96.50	78.50
Lite-HRNet	Lite-HRNet-18	68.01	92.77	73.10	72.00	94.00	76.00
Lite-HRNet	Lite-HRNet-30	67.23	96.64	73.01	70.55	97.00	75.00
HPFPE（Ours）	HRNet-W32	72.12	97.98	74.87	76.30	98.50	78.50
HPFPE（Ours）	HRNet-W48	72.91	96.90	78.13	76.65	97.00	80.50

Method	Backbone	AP/%	AP₅₀/%	AP₇₅/%	AR/%	AR₅₀/%	AR₇₅/%
CPM	CPM	67.26	94.42	71.60	71.85	95.50	76.00
SCNet	SCNet-50	71.62	96.82	75.35	75.20	97.50	78.50
SCNet	SCNet-101	71.67	96.82	77.94	75.60	97.50	81.00
Lite-HRNet	Lite-HRNet-18	69.96	97.76	72.71	73.60	98.00	76.50
Lite-HRNet	Lite-HRNet-30	70.01	93.64	75.72	73.75	95.00	78.00
HPFPE（Ours）	HRNet-W32	74.05	99.01	75.82	77.85	99.50	80.00
HPFPE（Ours）	HRNet-W48	74.12	97.89	81.99	77.60	98.00	84.00

Method	Backbone	Input size	AP/%	AP₅₀/%	AP₇₅/%	AR/%	AR₅₀/%	AR₇₅/%
HRNet+CBAM_front	HRNet-W32	256×192	71.85	95.90	78.60	75.80	96.50	81.50
	HRNet-W32	384×288	73.46	96.88	79.33	76.75	97.50	82.00
	HRNet-W48	256×192	71.37	96.93	76.45	75.15	97.50	80.50
	HRNet-W48	384×288	73.23	96.66	77.90	76.45	97.00	80.50
HRNet+CBAM_fuse	HRNet-W32	256×192	70.95	95.94	77.42	74.85	96.50	80.50
	HRNet-W32	384×288	71.93	95.78	75.41	76.10	96.50	78.50
	HRNet-W48	256×192	71.09	96.81	77.20	75.75	97.00	81.00
	HRNet-W48	384×288	69.61	96.96	73.95	73.05	97.00	76.50
HRNet+CBAM_stage	HRNet-W32	256×192	71.94	93.83	80.43	75.90	95.00	82.50
	HRNet-W32	384×288	74.02	95.94	80.49	77.40	96.50	82.50
	HRNet-W48	256×192	71.86	95.79	81.60	76.00	96.00	84.00
	HRNet-W48	384×288	73.40	97.85	78.87	76.75	98.00	81.00

Attention mechanism	Backbone	Input size	AP/%	AP₅₀/%	AP₇₅/%	AR/%	AR₅₀/%	AR₇₅/%
SE^［24］	HRNet-W32	256×192	71.53	97.89	77.95	76.30	98.00	81.00
	HRNet-W32	384×288	73.23	96.88	78.60	77.20	97.50	81.00
	HRNet-W48	256×192	71.55	97.95	81.97	74.95	98.50	84.00
	HRNet-W48	384×288	72.79	96.81	78.37	76.25	97.00	81.00
ECA^［25］	HRNet-W32	256×192	71.59	96.95	76.18	75.30	97.50	79.50
	HRNet-W32	384×288	72.49	97.89	79.66	77.35	98.00	83.00
	HRNet-W48	256×192	71.82	94.56	76.64	75.70	95.50	79.00
	HRNet-W48	384×288	72.79	96.81	78.37	76.25	97.00	81.00
CA^［26］	HRNet-W32	256×192	71.02	94.94	77.60	74.50	95.50	80.00
	HRNet-W32	384×288	73.03	97.95	77.97	75.75	98.00	80.00
	HRNet-W48	256×192	70.55	94.97	77.74	74.40	95.50	80.50
	HRNet-W48	384×288	72.82	97.86	77.39	76.75	98.00	80.00
LSKblock^［27］	HRNet-W32	256×192	71.71	97.89	77.40	75.35	98.00	80.00
	HRNet-W32	384×288	72.26	96.93	77.37	75.55	97.50	79.00
	HRNet-W48	256×192	71.83	95.85	78.14	75.75	96.50	81.50
	HRNet-W48	384×288	71.89	95.95	76.91	75.00	96.00	79.00
CBAM^［23］	HRNet-W32	256×192	71.94	93.83	80.43	75.90	95.00	82.50
	HRNet-W32	384×288	74.02	95.94	80.49	77.40	96.50	82.50
	HRNet-W48	256×192	71.86	95.79	81.60	76.00	96.00	84.00
	HRNet-W48	384×288	73.40	97.85	78.87	76.75	98.00	81.00

[1]	黎祖胜, 唐吉深, 匡迎春. 基于改进YOLOv10n的轻量化荔枝虫害小目标检测模型[J]. 智慧农业(中英文), 2025, 7(2): 146-159.
[2]	牛子昂, 裘正军. 基于改进YOLOv11-Pose的玉米植株骨架及表型参数提取方法[J]. 智慧农业(中英文), 2025, 7(2): 95-105.
[3]	吴六爱, 许雪珂. 基于改进YOLOv10n的轻量化番茄叶片病虫害检测方法[J]. 智慧农业(中英文), 2025, 7(1): 146-155.
[4]	权家璐, 陈雯柏, 王一群, 程佳璟, 刘亦隆. 基于GCN-BiGRU-STMHSA的农业干旱预测研究[J]. 智慧农业(中英文), 2025, 7(1): 156-164.
[5]	齐梓均, 牛当当, 吴华瑞, 张礼麟, 王仑峰, 张宏鸣. 基于双维信息与剪枝的中文猕猴桃文本命名实体识别方法[J]. 智慧农业(中英文), 2025, 7(1): 44-56.
[6]	芦碧波, 梁迪, 杨洁, 宋爱青, 皇甫尚卫. 基于改进ENet的复杂背景下山药叶片图像分割方法[J]. 智慧农业(中英文), 2024, 6(6): 109-120.
[7]	周秀珊, 文露婷, 介百飞, 郑海锋, 吴其琦, 李克讷, 梁军能, 黎一键, 文家燕, 江林源. 改进YOLOv11的水面膨化饲料颗粒图像实时检测算法[J]. 智慧农业(中英文), 2024, 6(6): 155-167.
[8]	胡程喜, 谭立新, 王文胤, 宋敏. 基于改进DeepLabV3+的轻量化茶叶嫩芽采摘点识别模型[J]. 智慧农业(中英文), 2024, 6(5): 119-127.
[9]	年悦, 赵凯旋, 姬江涛. 基于改进DeepLabCut模型的奶牛滑蹄检测方法[J]. 智慧农业(中英文), 2024, 6(5): 153-163.
[10]	李明煌, 苏力德, 张永, 宗哲英, 张顺. 基于改进YOLOv8n-pose和三维点云分析的蒙古马体尺自动测量方法[J]. 智慧农业(中英文), 2024, 6(4): 91-102.
[11]	翁智, 范琦, 郑志强. 基于多模态图像信息及改进实例分割网络的肉牛体尺自动测量方法[J]. 智慧农业(中英文), 2024, 6(4): 64-75.
[12]	范铭铄, 周平, 李淼, 李华龙, 刘先旺, 麻之润. 羊场自动导航喷药机器人设计与实验[J]. 智慧农业(中英文), 2024, 6(4): 103-115.
[13]	侯依廷, 饶元, 宋贺, 聂振君, 王坦, 何豪旭. 复杂大田场景下基于改进YOLOv8的小麦幼苗期叶片数快速检测方法[J]. 智慧农业(中英文), 2024, 6(4): 128-137.
[14]	翁智, 刘海鑫, 郑志强. CSD-YOLOv8s：基于无人机图像的密集小目标羊只检测模型[J]. 智慧农业(中英文), 2024, 6(4): 42-52.
[15]	王宇啸, 石源源, 陈招达, 吴珍芳, 蔡更元, 张素敏, 尹令. 猪三维点云体尺自动计算模型Pig Back Transformer[J]. 智慧农业(中英文), 2024, 6(4): 76-90.

Method	Backbone	Input size	AP/%	AP₅₀/%	AP₇₅/%	AR/%	AR₅₀/%	AR₇₅/%
DeepPose	ResNet-50	256×192	4.05	21.50	0.20	11.25	40.00	2.50
	ResNet-101	256×192	7.66	30.80	0.00	13.75	50.00	0.00
	ResNet-152	256×192	8.48	32.07	0.08	16.00	50.00	2.50
CPM	CPM	256×192	27.25	84.78	6.01	34.50	85.00	17.50
CPM	CPM	384×288	33.76	89.88	18.92	42.50	90.00	37.50
SCNet	SCNet-50	256×192	44.04	89.89	40.68	51.25	90.00	55.00
	SCNet-50	384×288	49.00	85.20	50.85	55.50	87.50	62.50
	SCNet-101	256×192	47.20	94.06	44.68	54.25	95.00	55.00
	SCNet-101	384×288	50.37	93.88	56.17	58.50	95.00	65.00
Lite-HRNet	Lite-HRNet-18	256×192	39.67	89.97	30.35	47.00	90.00	45.00
	Lite-HRNet-18	384×288	45.72	94.06	35.12	52.50	95.00	50.00
	Lite-HRNet-30	256×192	39.82	80.55	41.12	46.75	82.50	55.00
	Lite-HRNet-30	384×288	43.02	90.10	37.32	50.75	90.00	52.50
HRNet	HRNet-W32	256×192	47.63	92.03	40.77	55.50	92.50	55.00
	HRNet-W32	384×288	48.11	89.25	41.87	56.00	90.00	57.50
	HRNet-W48	256×192	47.25	87.13	42.85	52.50	87.50	55.00
	HRNet-W48	384×288	50.76	94.06	40.92	58.25	95.00	57.50
HPFPE（Ours）	HRNet-W32	256×192	47.88	92.03	33.28	54.25	92.50	50.00
	HRNet-W32	384×288	49.50	90.46	38.23	56.25	92.50	55.00
	HRNet-W48	256×192	48.54	84.91	44.25	55.50	87.50	57.50
	HRNet-W48	384×288	52.96	91.18	52.67	59.50	92.50	62.50