EventPointMesh: Human Mesh Recovery Solely From Event Point Clouds.

Area of Science:

• Computer Vision
• Neuromorphic Engineering
• Human Shape Analysis

Background:

• Traditional human mesh recovery (HMR) relies on intensity images, which are susceptible to lighting variations, privacy concerns, and high resource demands.
• Event cameras offer an alternative by capturing luminance changes at high speeds, independent of lighting conditions, but their use in HMR is underexplored.

Purpose of the Study:

• To investigate the feasibility of reconstructing human shape using solely event camera data.
• To develop a novel framework, EventPointMesh, for HMR that overcomes the limitations of intensity-based methods.

Main Methods:

• Proposed EventPointMesh framework treating event data as a 3D spatio-temporal point cloud.
• Implemented a coarse-to-fine pose feature extraction strategy to capture both global and local human body features.
• Grouped dispersed event points into segments corresponding to individual body parts for detailed analysis.

Main Results:

• Demonstrated accurate human mesh recovery using only sparse event data.
• EventPointMesh outperformed existing baseline methods in HMR tasks.
• The method effectively captures subtle human movements through combined global and local feature extraction.

Conclusions:

• Event camera data alone provides sufficient information for robust human mesh recovery.
• EventPointMesh offers a privacy-preserving, low-light-resilient, and efficient alternative for HMR.
• The proposed approach advances neuromorphic vision applications in human pose and shape estimation.