At IIT Madras, I investigated deep-learning-based shape skeletonization as a Young Research Fellow, benchmarking more than 10 approaches. I modeled midcurves and midsurfaces with signed distance fields under Prof. Ramanathan Muthuganapathy.
For DistillSkel, I distilled a U-Net from 31M to 31K parameters with under 5% F1 loss in skeletonization accuracy, and introduced a junction-preservation loss. I also led the in-house creation of a 2k+ shape dataset through augmentation, preprocessing and guided synthesis.
My work also includes graph learning for solar-flare classification, where combining GATs and LSTMs improved accuracy by more than 8% over non-graph baselines. In a multi-view 3D reconstruction project, I used 3D Gaussian Splatting to achieve 77.2% accuracy compared with 56.8%.
