Conferencia: "Learning Motion Patterns for Weakly-supervised Semantic Segmentation”.
Ponente: Karteek Alahari (INRIA Grenoble - Rhône-Alpes)
Organiza: Nicolás Guil Mata, Dpto. Arquitectura de Computadores.
Línea: Computación de Altas Prestaciones
Lugar: Sala de grados A. ETSI Informática.
Fecha y hora: 20/04/2018 a las 12:00


No es necesaria pre-inscripción; entrada libre hasta completar el aforo.


Contenido: Weakly-supervised learning for semantic segmentation has emerged as a core problem in the
context of deep learning frameworks. Despite this, their performance has remained significantly behind
that of fully-supervised approaches. One of the main issues is the inability of weakly-supervised methods
to accurately capture object boundaries. We address this problem through our framework for automatically
learning object contours from motion cues, relying on video-level labels and synthetic datasets. This talk
will present our contributions in this area over the past two years
In our initial work (presented at ECCV'16) we leverage unsupervised video segmentation for obtaining
object contours, and learn a semantic segmentation model from labels assigned to the entire video, in
contrast to pixel-level annotation. Inspired by the utility of these pre-computed motion cues, we turn to a
learning-based approach for estimating them directly from video, and propose the first CNN for motion
segmentation (presented at CVPR'17), which is later extended to a video object segmentation framework
by augmenting it with a visual memory module (presented at ICCV'17). Our recent work combines the two
models for estimating motion cues and segmentation into an end-to-end weakly-supervised semantic
segmentation model.

 

La asistencia a este acto está contemplada dentro de las Actividades Formativas para los doctorandos de
nuestro programa. Para que se considere como actividad formativa, además de la asistencia, deberá
subirse a DAD un resumen crítico realizado por el estudiante, que será validado por el tutor.