M.S. Candidate: Ömer Faruk Kürklü
Program: Multimedia Informatics
Date: 15.09.2026 / 10:00
Place: A-212

Abstract: We introduce PolyEcho, a permutation-invariant model that reconstructs a receiver-centered vector floor plan from any set of 4-16 irregularly positioned, position-tagged single-channel room impulse responses (RIRs). Unlike systems tied to a fixed sensor array, PolyEcho is trained using a different blind source constellation in every room and operates on randomly selected measurement subsets. A single checkpoint improves from 86.3% mean intersection over union (IoU, measured between the reconstructed room and the ground truth room) with four measurements to 96.2% with sixteen. Acquisition cost, spatial coverage and reconstruction accuracy can therefore be traded off at deployment without retraining. Training on random measurement subsets simultaneously enables variable-cardinality inference and acts as structured measurement dropout regularization. Receiver-relative source positions resolve the global planar gauge and provide coarse interior support, while acoustic evidence contributes a further 18.9 IoU points over positions alone at R=16. The model returns an ordered vector contour and is evaluated on a 240,000-room general-polygon benchmark that spans 4-8 corners and up to five reflex vertices, and includes certified first-order-invisible walls.