Here is a list of all my projects related to action recognition, medical imaging and data collection.
Particle Flow Fields (September 2011 – present)
“Particle Flow Fields” is a new video representation based on optical flow and particle advection. PFF outperforms optical flow and gradients. Particle Flow Fields also showed good performance on Kinect Gesture Recognition dataset in a “One-Shot-Learning” framework.
3D- Spatio Temporal Volumes (May 2011 – August 2011)
Investigated the performance of 3D gradients and PFF for action recognition task in aerial videos taken from an aircraft and performed extensive testing on aerial and rooftop videos in VIRAT dataset. The designed 3D-STHOG descriptor is incorporated into VIRAT DARPA system for testing.
Scene Conext Descriptor (October 2011 – May 2011)
Scene Context descriptor is proposed. Demonstrated improvement in accuracy on different action recognition datasets by using scene context information in conjunction to any motion descriptor. The datasets include UCF11, UCF50, HMDB51 and a sample from TRECVID dataset released as part of ALADDIN project.
KEYS System (March 2009 – September 2009)
KEYS system is developed in C++ based on feature-trees algorithm for real-time action recognition. New action categories can be added to the system on the fly. Demonstrated KEYS action recognition system live at ICCV2009 conference held in Japan.
Feature-tree (January 2009 – June 2009)
Developed an algorithm to perform multiple action recognition, incremental action recognition and action localization using feature-trees
Video Image Retrieval and Analysis Tool (VIRAT) project (Aug 2008 – Feb 2012)
This project is sponsored by DARPA to do action recognition in aerial videos. Action recognition algorithms were tested on APHill dataset. Two action recognition algorithms were implemented in c++ and integrated in into the VIRAT DARPA system developed by Lockheed Martin and Kitware
Brain tumor detection and segmentation in MRI images (May 2009 – Nov 2011)
This project is sponsored by National Institutes of Health (NIH). The task is to do brain tumor detection and segmentation in MRI images. We developed an algorithm using texture features and multi MRI modalities to generate confidence surface and used it to perform brain tumor detection and segmentation.
UCF-50 Action Recognition Dataset (July 2010 – October 2010)
UCF-50 is one of the biggest action recognition datasets. This dataset has about 6676 videos taken from YouTube and has 50 different action categories.
UCF-AerialRooftopGround (June 2010 – May2011)
UCF-ARG is a unique action recognition dataset, where the action is simultaneously captured from a ground camera, a rooftop camera at 100 feet and a remotely controlled aerial camera mounted on a helium balloon.
UCF-iPhone Action Dataset (January 2011 – March 2011)
UCF-iPhone dataset was generated by recording data from an Inertial Measurement Unit (IMU) on an Apple iPhone 4 for 9 different actions performed by different actors.