Ankur Handa

handa(dot)ankur(at)gmail(dot)com

Scholar

Hello World! I obtained my PhD in the Department of Computing under the supervision of Dr. Andrew Davison. I work as a post-doctoral research assistant at University of Cambridge.

Publications

Real-Time Camera Tracking: When is High Frame-Rate Best?Ankur Handa, Richard A. Newcombe, Adrien Angeli, Andrew J. Davison [pdf] [poster] [video] [Dataset]
ECCV 2012.
[Abstract]
Higher frame-rates promise better tracking of rapid motion, but advanced real-time vision systems rarely exceed the standard 10-60Hz range, arguing that the computation required would be too great. Actually, increasing frame-rate is mitigated by reduced computational cost per frame in trackers which take advantage of prediction. Additionally, when we consider the physics of image formation, high frame-rate implies that the upper bound on shutter time is reduced, leading to less motion blur but more noise. So, putting these factors together, how are application-dependent performance requirements of accuracy, robustness and computational cost optimised as frame-rate varies? Using 3D camera tracking as our test problem, and analysing a fundamental dense whole image alignment approach, we open up a route to a systematic investigation via the careful synthesis of photorealistic video using ray-tracing of a detailed 3D scene, experimentally obtained photometric response and noise models, and rapid camera motions. Our multi-frame-rate, multi-resolution, multi-light-level dataset is based on tens of thousands of hours of CPU rendering time. Our experiments lead to quantitative conclusions about frame-rate selection and highlight the crucial role of full consideration of physical image formation in pushing tracking performance.


Scalable Active Matching. Ankur Handa, Margarita Chli, Hauke Strasdat, Andrew J. Davison [pdf] [poster] [old video] [new video]
CVPR 2010.
[Abstract]
Matching image features given general priors on their locations and correlations among them, can be efficiently solved guided by "do computation only which is really required" paradigm. Active Matching works exactly in this manner exploiting these priors to decide the best choice available at hand at each step that would yield the maximum information gain. However, this sequentially guided approach expends a considerable effort in deciding the importance of choices thereby limiting the scalability of the approach to only a conservative number of features. In this paper, borrowing the theory and underlying principles from Active Matching, we relax the general structure of the problem it maintains to conservative approximations with only a slight accuracy penance while increasing the efficiency by high proportions. One approach, CLAM (Chow-Liu Active Matching), reduces the dense graph maintained by Active Matching to a maximum-spanning tree while the other, SubAM (Subset Active Matching) works by building a tree of subsets(clusters) of features. In essence, both approaches thwart the unmanageably explosive computation of a canonical dense bayesian graph structure maintained by Active Matching algorithm.

Technical Reports

Simplified Jacobians in 6-DoF Camera Tracking. Ankur Handa [pdf]
Aug 2014

Applications of Legendre-Fenchel Transformation. Ankur Handa, Richard A. Newcombe, Adrien Angeli, Andrew J. Davison [pdf][bibtex]
Sept 2011.
This report attempts to explain some useful details and minutiae associated with Legendre-Fenchel transformation. It is not a deep investigation into the topic but it is useful if you are getting started with variational methods and wanting to implement standard energy based minimisation methods e.g. Image denoising, Optical Flow, Superresolution, Depth-map estimation or camera tracking. Most of the details can be easily adapted to other new energy formulations based on variatonal methods involving primarily the notorious L1 norm. Together with Richard, Adrien and Andy, I spent some time this summer trying to understand variational based theory made popular by GPU4Vision. All these various methods explained in the report are extremely amenable to GPUs as well as very trivial to implement. Most of the implementations would be made available very shortly. We would also like to thank Thomas Pock, Hauke Strasdat as well as Luis Pizarro for helpful discussions.
ROF CUDA code

Past
Slightly ancient and odd looking home-page. [Old Home Page]

Other Links