Personal tools
You are here: Home Publications Temporally Consistent Disparity Maps

Temporally Consistent Disparity Maps Using CRFs with Fast 4D Filtering

Asian Conference on Pattern Recognition (ACPR), 2015, oral presentation

Sivash Bigdeli
Gregor Budweiser
Matthias Zwicker
University of Bern
3D Impact Media
University of Bern

Abstract

State of the art methods for disparity estimation achieve good results for single stereo frames, but temporal coherence in stereo videos is often neglected. In this paper we present a method to compute temporally coherent disparity maps. We define an energy over whole stereo sequences and optimize their Conditional Random Field (CRF) distributions using mean-field approximation. We introduce novel terms for smoothness and consistency between the left and right views, and perform CRF optimization by fast, iterative spatio-temporal filtering with linear complexity in the total number of pixels. Our results rank among the state of the art while having significantly less flickering artifacts in stereo sequences.

Materials

publication, supplementary material

Document Actions