Core Concepts
提案されたDOR3D-Netは、3D手の姿勢推定を密な序数回帰問題として再構築し、SOTA手法に比べて優れた性能を示す。
Abstract
I. Introduction
Depth-based 3D hand pose estimation is crucial in human-machine interaction.
Dense regression methods have gained attention for their accuracy and low computational burden.
II. Related Work
Depth image-based hand pose estimation methods categorized into regression-based and detection-based methods.
III. Methods
Feature extractor module uses a transformer-based approach to capture long-range relationships.
Dense ordinal regression module predicts probability maps with ordinal constraints.
IV. Experiments
Datasets used include HANDS2017, MSRA, ICVL, and NYU for evaluation.
DOR3D-Net outperforms SOTA methods in terms of mean error and success frames percentage.
Stats
提案されたネットワークは、HANDS2017、MSRA、ICVL、およびNYUデータセットで最先端の性能を達成しました。