M. Pawan Kumar
 
 

HOME

RESEARCH

PUBLICATIONS

GROUP

TALKS

TEACHING

CV

 

 

 

 

 

 

LEARNING TO RANK USING HIGH-ORDER INFORMATION

P. Dokania, A. Behl, C. V. Jawahar and M. Pawan Kumar

We consider the problem of using high-order information (for example, persons in the same image tend to perform the same action) to improve the accuracy of ranking (specifically, average precision). We develop two learning frameworks. The high-order binary SVM (HOB-SVM) optimizes a convex upper bound of the surrogate 0-1 loss function. The high-order average precision SVM (HOAP-SVM) optimizes a difference-of-convex upper bound on the average precision loss function.


OPTIMIZING AVERAGE PRECISION USING WEAKLY SUPERVISED DATA

A. Behl, C. V. Jawahar and M. Pawan Kumar

We develop a novel latent AP-SVM that minimizes a carefully designed upper bound on the AP-based loss function over a weakly supervised dataset. Our approach is based on the hypothesis that in the challenging setting of weakly supervised learning, it becomes crucial to optimize the right accuracy measure. Using publicly available datasets, we demonstrate the advantage of our approach over standard loss-based binary classifiers on challenging problems in computer vision.


SELF-PACED LEARNING FOR LATENT VARIABLE MODELS

M. Pawan Kumar, B. Packer and D. Koller

We develop an accurate iterative algorithm for learning the parameters of a latent variable model such as latent structural SVM. Our approach uses the intuition that the learner should be presented with the training samples in a meaningful order: easy samples first, hard samples later. At each iteration, our approach simultaneously chooses the easy samples and updates the parameters.


REGION SELECTION FOR SCENE UNDERSTANDING

M. Pawan Kumar and D. Koller

We consider the problem of simultaneously dividing an image into coherent regions and assigning labels to regions using a global energy function. We form a large dictionary of putative regions using bottom-up over-segmentation techniques and formulate the problem of selecting the regions and their labels as an integer program. We provide an efficient dual decomposition method to solve an accurate linear programming relaxation of the integer program.


IMPROVED MOVES FOR TRUNCATED CONVEX MODELS

M. Pawan Kumar and P. Torr

We develop a new st-MINCUT based move-making method for MAP estimation of discrete MRFs with arbitrary unary potentials and truncated convex pairwise potentials. We prove that our method provides the best known multiplicative bounds (same as the bounds obtained by solving the standard linear programming relaxation followed by randomized rounding) for these problems in polynomial time. We demonstrate the efficacy of our approach using synthetic and real data experiments.


ANALYSIS OF CONVEX RELAXATIONS FOR MAP ESTIMATION

M. Pawan Kumar, V. Kolmogorov and P. Torr

We analyze several convex relaxations for MAP estimation of discrete MRFs. We show that the standard linear programming relaxation dominates (provides a tighter approximation than) a large class of quadratic programming and second order cone programming relaxations. Our analysis leads to new second order cone programming relaxations that are tighter than the linear programming relaxation.


OBJECT CATEGORY SPECIFIC SEGMENTATION

M. Pawan Kumar, P. Torr and A. Zisserman

Given an image containing an instance of a known object category, we obtain an accurate, object-like segmentation automatically. We match a parts-based object category model to the image. Each sample of the model provides cues about the shape of the particular instance, which are incorporated in a global energy function. The segmentation is obtained by minimizing the energy using a single st-MINCUT.


LEARNING LAYERED MOTION SEGMENTATIONS

M. Pawan Kumar, P. Torr and A. Zisserman

Given a video we learn a layered representation of the scene for motion segmentation in an unsupervised manner. The layered representation consists of the shape and appearance of the various rigidly moving segments in the scene, their occlusion ordering as well as their framewise transformations. These parameters are estimated from a video by minimizing a global energy function using block coordinate descent.