ITE Technical Group Submission System
Conference Paper's Information
Online Proceedings
[Sign in]
 Go Top Page Go Previous   [Japanese] / [English] 

Paper Abstract and Keywords
Presentation 2024-02-19 13:45
Efficient Human Pose and Shape Estimation using Decomposed Manhattan Self-Attention
Yushan Wang, Botao Zhang (TMU), Shuhei Tarashima (NTT Com), Norio Tagawa (TMU)
Abstract (in Japanese) (See Japanese page) 
(in English) HMR2.0, a high performance human pose and shape estimation algorithm, leverages ViT as its backbone and uses pretrained weights that has learned spatial relationships, leads to high number of parameters and complexity. Our goal is to significantly reduce both parameters and model complexity while preserving the model's expressive capability to a considerable extent. We replace the ViT backbone with spatial decay matrix and proposed decomposed manhattan-attention based architecture, which characterized by its linear complexity. We mix the typical datasets for training with different weights as in HMR2.0, i.e., Human3.6M 0.1, MPI-INF3DHP 0.02, COCO 0.2, MPII 0.1, InstaVariety 0.2, AVA 0.19 and AI Challenger 0.19. We compare the parameters and FLOPs between HMR2.0 and our proposed Decomposed Manhattan Self-Attention based linear complexity structure. Experimental results show that we reduce FLOPs from 242.1G to 17.5G. In terms of qualitative comparison, the adoption of linear complexity led to inferior results compared to the HMR2.0. This outcome was anticipated as, in HMR2.0, to attain optimal results, pre-training weights based on ImageNet were initially employed. However, due to modifications of linear complexity in our network structure, the use of the original pre-trained weights became impractical, necessitating a complete restart of training from scratch.
Keyword (in Japanese) (See Japanese page) 
(in English) Pose and Shape Estimation / ViT / HMR2.0 / Linear Complexity / / / /  
Reference Info. ITE Tech. Rep., vol. 48, no. 6, ME2024-25, pp. 44-48, Feb. 2024.
Paper # ME2024-25 
Date of Issue 2024-02-12 (MMS, ME, AIT) 
ISSN Online edition: ISSN 2424-1970
Download PDF

Conference Information
Committee IEICE-ITS IEICE-IE ME AIT MMS  
Conference Date 2024-02-19 - 2024-02-20 
Place (in Japanese) (See Japanese page) 
Place (in English) Hokkaido Univ. 
Topics (in Japanese) (See Japanese page) 
Topics (in English) Image Processing, etc. 
Paper Information
Registration To ME 
Conference Code 2024-02-ITS-IE-ME-AIT-MMS 
Language English 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Efficient Human Pose and Shape Estimation using Decomposed Manhattan Self-Attention 
Sub Title (in English)  
Keyword(1) Pose and Shape Estimation  
Keyword(2) ViT  
Keyword(3) HMR2.0  
Keyword(4) Linear Complexity  
Keyword(5)  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Yushan Wang  
1st Author's Affiliation Tokyo Metropolitan University (TMU)
2nd Author's Name Botao Zhang  
2nd Author's Affiliation Tokyo Metropolitan University (TMU)
3rd Author's Name Shuhei Tarashima  
3rd Author's Affiliation NTT Communications Corporation (NTT Com)
4th Author's Name Norio Tagawa  
4th Author's Affiliation Tokyo Metropolitan University (TMU)
5th Author's Name  
5th Author's Affiliation ()
6th Author's Name  
6th Author's Affiliation ()
7th Author's Name  
7th Author's Affiliation ()
8th Author's Name  
8th Author's Affiliation ()
9th Author's Name  
9th Author's Affiliation ()
10th Author's Name  
10th Author's Affiliation ()
11th Author's Name  
11th Author's Affiliation ()
12th Author's Name  
12th Author's Affiliation ()
13th Author's Name  
13th Author's Affiliation ()
14th Author's Name  
14th Author's Affiliation ()
15th Author's Name  
15th Author's Affiliation ()
16th Author's Name  
16th Author's Affiliation ()
17th Author's Name  
17th Author's Affiliation ()
18th Author's Name  
18th Author's Affiliation ()
19th Author's Name  
19th Author's Affiliation ()
20th Author's Name  
20th Author's Affiliation ()
Speaker Author-1 
Date Time 2024-02-19 13:45:00 
Presentation Time 15 minutes 
Registration for ME 
Paper # MMS2024-9, ME2024-25, AIT2024-9 
Volume (vol) vol.48 
Number (no) no.6 
Page pp.44-48 
#Pages
Date of Issue 2024-02-12 (MMS, ME, AIT) 


[Return to Top Page]

[Return to ITE Web Page]


The Institute of Image Information and Television Engineers (ITE), Japan