Work

M.Tech (Master's) Project - (Holi'13 - Holi'14)

My M.Tech Project (MTP) was in the field of Video Activity Recognition under the guidance of Dr. Parag Singla.

Title of my project was Incorporating Object and People Information to Improve Video Activity Recognition. We used Markov Logic Networks(MLNs) to capture the relationship between 'Activity' and 'Object & People' information. MLNs use domain knowledge in the form of First Order Logic (FOL) and probabilities in forms of weighted rules in FOL. For example, existence of a car improves probability of activity 'drive_car' over 'eating'. And, presence of a two people increases probability of activity 'shake_hands' over 'drive_car'. Weighted FOL formula can help in such predictions.

Our model uses FOL formulae such as :

and, using MLN, we learn weights for these formulae. A highly intuitive rule will get a higher weight as compared to a counter intuitive rule.

Project Timeline / Progress :

• 1. (June'14) Incorporating semantic information using MLNs is done. Improved results are included in the thesis (866KB).

  2. (March'14) I have implemented activity and object recognition. Now working on integrating MLNs.

  3. (December'13) I have implemented activity recognition. Now working on object recognition.

  4. (April'13) I am partially implementing the paper "Improving Video Activity Recognition using Object Recognition and Text Mining" by Motwani et al.

  5. (March'13) I am doing literature survey in the field of [Human] activity detection.

What novel do we add by using MLN? -

• 1. Domain knowledge - Existence of certain object may increase confidence of activity being present/absent in the video clip. For example, presence of a dining table will increase confidence of eating and decrease confidence of driving a car.

  2. Inherent probabilistic behaviour of inference - Standard FOL encapsulates knowledge in hard clauses. The weights of each clause in MLN can encapsulate probabilistic behaviour. If a clause is quite unlikely, then its weight would be low as compared to highly probable clause.

M.Tech Thesis : Incorporating Object and People Information to Improve Video Activity Recognition (866KB)

Bachelor's (B.Tech) Project - 2010-11

My final year B.Tech project was in the field of compilers (GCC). I, along with two other students ( Pratik Patre and Waman Virgaonkar ) worked on heap reference analysis under the guidance from Prof. C. S. Moghe from VNIT and Prof. U. P. Khedker from IIT, Bombay. Main aim of our project was to improve conventional garbage collection which after improvement, will work on liveness of memory location rather than just the reachability.

• • B.Tech Thesis : Heap Reference Analysis and Its Implementation in GCC (2.0MB)

  • The GCC Pass(hra_patch_vnit.tar : a tar file of the patch)