Campus Event Calendar

Event Entry

What and Who

Joint Co-localisation and Co-segmentation

Abhishek Sharma
Inria, Saclay
AG 1, AG 2, AG 3, AG 4, AG 5, RG1, SWS, MMCI  
MPI Audience

Date, Time and Location

Monday, 11 May 2015
45 Minutes
E1 4


Co-segmentation is the problem of dividing a set of images into regions corresponding to k classes. Previous attempts to solve this problem can be broadly classified into two paradigms: Discriminative [1] and similarity based approach. Discriminative framework[1] do not ímpose any constraint that regions with the same label (eg. foreground) should have similar visual feature distributions. Moreover, the resulting objective function is optimized with a SDP solver, which is not scalable for more than (say) 50 images. Building on the work of [1], we include a foreground model and a linear prior and optimise the resulting function with QP solver, making the current framework suitable to large scale.

Given the successful trend of using object proposals in object detection, we extend the idea for co-segmentation. Intuition being that better spatial support will help discriminative clustering to focus on rich set of features. In the second part, we show how combining Co-localisation [2] with co-segmentation achieves very competitive results on three datasets.


Connie Balzert
0681 9325 2000
--email hidden
passcode not visible
logged in users only

Connie Balzert, 05/06/2015 13:51
Connie Balzert, 05/06/2015 13:50 -- Created document.