- Built an end-to-end basketball game analysis system that can identify team and number of players, track movements of players, analyze body postures and team tactics, transform perspective and map it onto a tactical board using machine learning
- Surveyed latest papers and adapted a different framework to resolve id-switching problem in object tracking (BoT-SORT)
- Implemented finest architecture by integrating two models into one and increased performance by 30% (Multi-task SE-ResNet)
- Experimented different image processing methods and increased team classification accuracy by 30% (Densepose)
- Compressed and deployed deep learning models on edge machines (Jetson Nano), increased speed by 300%