This week, I worked on finishing some UI fixes and integrating everything together. First, I finished the nutrition information popup:
Then, Samuel and I spray painted the platform for better visibility. The computer vision algorithm improved drastically with the nicer white background.
Next, I set up the Jetson for the computer vision processing. We had a lot of issues with performance, but I discovered that when I disabled all of the GUI processes, the performance improved slightly. We wanted to improve performance using Tensor RT, but I found it very hard to compile (resource limited and arm processor) and no precompiled binaries were available. Fortunately, disabling the GUI was enough to improve performance, but it appears that there is a bug with torchscript that causes the performance to be extremely slow on the first detection. After “warming up” it seems to perform much better, so we will account for this in the code. My hypothesis is that the CPU needs to optimize what goes in the swapfile since it has to use about 1 GB of swap to run the process, which is significantly slower than RAM.
Finally, I tested the performance of the algorithm and we worked out a few bugs on the API and made sure that we are in good shape for the demo. I also collected detection times for our performance evaluation. I then worked on putting all this data into the final presentation.