SP-HRNet: single-path neural architecture search for human pose estimation
Abstract In this study, focus on the task of human pose estimation. High-resolution networks (HRNet), while adept at continuously integrating multi-resolution information, tend to compel the network to assimilate redundant information, thereby amplifying computational complexity. The single-path one...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-06-01
|
| Series: | Discover Applied Sciences |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s42452-025-06846-0 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract In this study, focus on the task of human pose estimation. High-resolution networks (HRNet), while adept at continuously integrating multi-resolution information, tend to compel the network to assimilate redundant information, thereby amplifying computational complexity. The single-path one-shot Neural Architecture Search (NAS) method, an efficacious solution to mitigate this issue. This research introduces a novel NAS method, denoted as SP-HRNet, based on Single-Path One-Step architecture. SP-HRNet meticulously selects the most suitable operator within a high-resolution network and identifies the optimal deep fusion construction methodology. Within SP-HRNet, we propose a hybrid search space incorporating an enhanced search strategy. Unlike the original high-resolution network, instead of modifying the network branches with a fixed decrement or increment, our network structure will be free to choose the most appropriate one among multiple branches. For this purpose, we devise two distinct search spaces labeled as Mixed Operator and Depth Fusion Options and formulate a novel search strategy tailored to this space. A myriad of experiments substantiates the effectiveness of this method. Our proposed network attains outstanding performance on the COCO human pose estimation dataset, achieving heightened accuracy while concurrently reducing the number of model parameters. |
|---|---|
| ISSN: | 3004-9261 |