Mengdi Shi Academic Portfolio
← Publications

Publication

PICFormer: Perception-Inference-Consistency Loop for Occluded 3D Pose Estimation

A closed-loop feedback Transformer architecture for 3D pose estimation under severe occlusion, featuring Visibility-aware Feature Modulation (VFM) and Gated Pyramid Attention (GPA) modules.

Date
January 1, 2026
Venue
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP, CCF-B)
Paper

This paper addresses the challenging problem of 3D pose estimation under severe occlusion. We propose PICFormer, the first closed-loop feedback Transformer architecture that enables dynamic information correction from downstream inference to upstream perception.

Key Contributions:

  • Designed and implemented Visibility-aware Feature Modulation (VFM) module
  • Developed Gated Pyramid Attention (GPA) mechanism
  • Achieved 2.9mm MPJPE reduction on 3DPW-Occ benchmark, outperforming all contemporary SOTA methods

Code: https://github.com/HKCyber20/PICFormer

Mengdi Shi, Yifeng Wang, Yi Zhao. (2026). "PICFormer: Perception-Inference-Consistency Loop for Occluded 3D Pose Estimation." ICASSP 2026.