3D Motion Estimation of Volumetric Deformable Objects from RGB-D Images Synthetically Generated by a Multi-camera System

Estimating deformation in volumetric objects, particularly when occluded, is a pressing challenge in computer vision. We present McDeforms, a novel dataset synthesized from a multi-camera system in PyBullet, simulating three scenarios of volumetric object deformations. Alongside RGB-D images, our dataset provides the object’s 3D coordinates ground truth and camera specifications. We explore McDeforms’ potentiality by evaluating two scene flow methods, Coherent Point Drift (CPD) and RAFT-3D, both of which competently estimate 3D flow across our simulations.