We faced several challenges while building our project. The most notable was converting the model to Intel's inference engine. We started by converting an MXNet model using the python converter but ran into several errors about missing axis attribute forconcat layer. After resolving these errors we were stuck with Test failed: This sample accepts networks having only one output, which meant we needed to restructure the final output layer or skip running on the GPU. FIX in version 1.2.2 - AWS Support Forums
Accomplishments that we're proud of
There were many times where we felt the project was a lost cause. Between the model not optimizing and interpreting the results, it was a bumpy road to a working demo. The first time seeing the live project feed with the pose map overlaying was finally the point where we knew we could finish what we started.
This project is not to be used for commercial use and you must follow the original license.
In the short term we would like to enable multiple pose detection to allow a group of players to play off the same device. The model is already setup to do this but we need to implement this into the game, as for a demo it was easier to setup for one player.
In the long term we envision a training/classification loop where, at random times, the system will ask the player to make a new action. We could then take the generated pose and update our classification weights.
Thinking about the larger picture, this pose estimation and classification has many applications from monitoring children, detecting shoplifting, other types of games like dancing, or maybe even in sport training.