ESCAPE: Equivariant Shape Completion via Anchor Point Encoding

TL;DR: Encode every point by its distances to a small set of anchor points so geometry is expressed in a rotation-equivariant form; a transformer maps partial observations to completed shapes without a separate pose-estimation stage.

Burak Bekci1 Mahdi Saleh1 Federico Tombari1,2 Nassir Navab1
1Technical University of Munich 2Google
CVPR 2025
Preprint Paper (PDF) Code Inference notebook
ESCAPE teaser — partial point clouds completed under varying rotations

Partial scans are completed consistently even when objects appear under arbitrary rotations and translations: distance-to-anchor encodings keep the representation aligned with 3D Euclidean symmetry, so the network reasons about shape rather than a fixed canonical pose.

ESCAPE model architecture

A transformer encodes and decodes anchor-distance features; predicted encodings are refined via optimization to recover dense completed point sets. The pipeline follows standard completion benchmarks (e.g., PCN, KITTI, OmniObject3D) with released training and evaluation scripts in the repository.

Abstract

ESCAPE (Equivariant Shape Completion via Anchor Point Encoding) targets rotation-equivariant shape completion from partial point clouds. The method selects anchor points on the shape and represents each point by its distances to all anchors, yielding a geometry-aware encoding that transforms consistently under rotations. A transformer architecture models these distance features end-to-end; an optimization step then maps predicted encodings back to completed shapes. Experiments show robust reconstructions across arbitrary rigid transforms without auxiliary pose-estimation modules, with code, pretrained models, and dataset pointers available in the project repository.

BibTeX

@InProceedings{Bekci_2025_CVPR,
  author    = {Bekci, Burak and Navab, Nassir and Tombari, Federico and Saleh, Mahdi},
  title     = {ESCAPE: Equivariant Shape Completion via Anchor Point Encoding},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2025},
  pages     = {6480--6489}
}