Abstract
Fitting a body to a 3D clothed human point cloud is a common yet challengingtask. Traditional optimization-based approaches use multi-stage pipelines thatare sensitive to pose initialization, while recent learning-based methods oftenstruggle with generalization across diverse poses and garment types. We proposeEquivariant Tightness Fitting for Clothed Humans, or ETCH, a novel pipelinethat estimates cloth-to-body surface mapping through locally approximate SE(3)equivariance, encoding tightness as displacement vectors from the cloth surfaceto the underlying body. Following this mapping, pose-invariant body featuresregress sparse body markers, simplifying clothed human fitting into aninner-body marker fitting task. Extensive experiments on CAPE and 4D-Dress showthat ETCH significantly outperforms state-of-the-art methods -- bothtightness-agnostic and tightness-aware -- in body fitting accuracy on looseclothing (16.7% ~ 69.5%) and shape accuracy (average 49.9%). Our equivarianttightness design can even reduce directional errors by (67.2% ~ 89.8%) inone-shot (or out-of-distribution) settings. Qualitative results demonstratestrong generalization of ETCH, regardless of challenging poses, unseen shapes,loose clothing, and non-rigid dynamics. We will release the code and modelssoon for research purposes at https://boqian-li.github.io/ETCH/.