Ph.D. Dissertation Defense: Rajeev Ranjan

Tuesday, January 15, 2019
11:00 a.m.
AVW 2328
Maria Hoo
301 405 3681
mch@umd.edu

ANNOUNCEMENT: Ph.D. Dissertation Defense

 
Name: Rajeev Ranjan
 
Advisory Committee:
Professor Rama Chellappa, Chair/Advisor
Professor David Jacobs
Professor Behtash Babadi
Professor Abhinav Shrivastava
Dr. Carlos D. Castillo
 
Day/time: Tuesday, January 15, 2019 at 11:00 am
 
Location: AVW 2328
 
Title: Advanced Regularizations in Deep Networks with applications in Face Analysis
 
Abstract:
 
Face Analysis in unconstrained settings is a challenging problem due to large variations in illumination, pose, expression, resolution, and appearance of a face. Though deep convolutional neural networks (DCNNs) have outperformed traditional methods on face detection, keypoints localization and face recognition, their performance can be enhanced using advanced regularization techniques such as multi-task learning (MTL) and feature compactness. 
 
The first part of the talk presents an algorithm for simultaneous face detection, keypoints localization, pose estimation and gender recognition using DCNNs. The proposed method, called HyperFace, fuses the intermediate layers of a DCNN followed by an MTL algorithm that operates on the fused features. This approach is extended to incorporate additional tasks of face verification, age estimation, and smile detection, in All-In-One Face. Both these methods exploit the synergy among the face related tasks which enhances the performance of the individual tasks. 
 
The second part of the talk presents a new loss function, called Crystal loss, to train a face recognition system. The loss adds a compactness constraint to the feature descriptors which enforces them to lie on a hypersphere manifold. It maximizes the inter-class distance and minimizes the intra-class distance in the cosine space which significantly boosts the performance of face verification. Additionally, the compactness constraint, when extended to the features of the convolutional layers, makes the network robust to multiple adversarial attacks.
 

Audience: Graduate  Faculty 

 

November 2019

SU MO TU WE TH FR SA
27 28 29 30 31 1 2
3 4 5 6 7 8 9
10 11 12 13 14 15 16
17 18 19 20 21 22 23
24 25 26 27 28 29 30
1 2 3 4 5 6 7
Submit an Event