Ph.D. Dissertation Defense: Wentao Luan

Tuesday, December 6, 2016
3:30 p.m.
Room 2328, AVW Bldg
Maria Hoo
301 405 3681
mch@umd.edu


ANNOUNCEMENT: Ph.D. Dissertation Defense

 
NAME: Wentao Luan

Committee:
Professor John S. Baras, Chair
Professor Cornelia Fermuller
Professor Yu Chen
Professor Behtash Babadi
Professor Yiannis Aloimonos, Dean's Representative

DATE/Time: Tuesday, Dec 06, 2016 at 3:30pm

PLACE: Room 2328, AVW Bldg
 

TITLE: Active Attention for Target Detection and Recognition in Robotics Vision

 
ABSTRACT:
Efficient and reliable target detection and recognition are among the fundamental problems in robot's system. Our work focuses on the problems in building a practical vision system for robotics and introduces attention mechanism into various levels of vision pipeline. Specifically, we study four situations. For single image processing, we study how to detect the target efficiently when only a detector(binary classifier) is available. The detection is then formulated as a sampling process. We exploit classifier's response score pattern to actively search the target in the image space and experiments demonstrate the efficiency of our method in reducing the number of classifications.
 
Still take a single image as input, but allowed to observe the target before, we explore on how to locate the target efficiently using the knowledge about the object. We propose to introduce recognition early into the traditional candidate proposal process. The target is described as a set of template graphs over the segmented target object surfaces, and detection becomes finding subgraphs from the input scene graph that match the template graphs. To focus on the target area efficiently, we propose a greedy constraints checking strategy and prove that it has a bounded performance concerning the optimal checking sequence.
 
On the control level, we investigate on the next viewpoint decision problem if the current observing quality would not satisfy the task. Our viewpoint selection module not only considers the viewing condition constraints for vision algorithms but also incorporates the low-level robot kinematics to guarantee the reachability of the desired viewpoint. By selecting viewpoints fast using a linear time cost score function, the system can deliver smooth user interaction experience. Additionally, we provide a learning from human demonstration method to obtain the score function parameters that better serves task's preference.
 
On information fusion level, we consider a problem in combining attribute classifiers' results to get a reliable output. For an active agent like a robot that can observe multiple times in the testing time, we propose to use two thresholds, one for high-precision prediction for positive class and the other for high precision negative class. Environment factor is also integrated into the decision making considering its influence to each component classifier. We prove our fusion framework's asymptotic correctness under certain assumptions on the attribute classifier and sufficient randomness of the input data.
 
 

Audience: Graduate  Faculty 

remind we with google calendar

 

March 2024

SU MO TU WE TH FR SA
25 26 27 28 29 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
31 1 2 3 4 5 6
Submit an Event