At Visist.ai, I fixed data and preprocessing issues in a video-based stroke-recognition pipeline, raising accuracy from 71% to 82%. I also adapted the BST paper into BST-HC, a hit-centric transformer that trained 8x faster, and built a forehand/backhand classifier from match footage.
In my ongoing M.Tech thesis at VISTA Lab, RBCCPS, IISc Bangalore, I’m developing methods to locate fixed CCTV cameras by matching their frames to Google Street View. My work includes a 3D approach that ran 3.8x faster than the original code and a 2D alternative that estimates heading 33–45x faster than the 3D baseline.
