My experiementation with BlazePose deep learning algorithm

This is the work I did as part of my engagement with Aerobi Inc as Machine Learning Developer. We are trying to implement a fast pose estimation algorithm that can run on an edge device such as mobile phone. The work done is part of MVP product.

The BlazePose deep pipeline has two portions. One portion is heatmap generation and second portion is regression pipeline to accurately predict key points. I was still working on improving accuracy of heatmaps to be generated. Once we have fairly accurate heatmaps, then move on to the regression pipeline.

Due to dynamic nature of start up environment, the goal changed and the project needed to be stopped in between.

I did lot of experimentation and hyper parameter tuning to reach closer to getting accurate heatmaps. I did get some limited success as well. Also wrote few custom loss functions as well as part of experimentation. Some early results are available to see in Results 20210226.zip file.