Semantic Labeling with Deep Learning and Surface Fitting

Giovedì 27 luglio alle 16, nell'aula "Archimede" del dipartimento di Matematica e Informatica (viale Andrea Doria, 6 - Catania), il dott. Pietro Zanuttigh, ricercatore al Dipartimento di Ingegneria dell'Informazione dell'Università di Padova, tiene il seminario dal titolo "Semantic Labeling with Deep Learning and Surface Fitting".

Attualmente, il prof. Zanuttigh lavora al Laboratorio di Tecnologia e Telecomunicazioni multimediali e la sua attività di ricerca riguarda principalmente l'acquisizione, compressione ed elaborazione di dati tridimensionali.

Abstract. The talk will address the exploitation of surface fitting clues inside deep learning classification frameworks. The first considered application is semantic segmentation of joint color and depth representations. 

The method starts with an initial over-segmentation based on spectral clustering followed by an iterative region merging procedure. The algorithm exploits a NURBS surface fitting scheme on the segments in order to understand if the selected couples correspond to a single surface and to build a set of additional descriptors for the neural network based on curvature information and fitting error. The color, depth and surface data is also fed to a Convolutional Neural Network (CNN) thus producing a per-pixel descriptor vector for each scene sample that is used both for semantic labeling and to compute a similarity metric that drives the merging process. The comparison with state-of-the-art methods shows how the proposed method provides an accurate and reliable segmentation and labeling of the scene. Finally the application of a variation of the approach for 3D objects classification will also be shown.

(27 luglio 2017)