The beginning of the week I tested the yolo balloon detection at a long range (20ft and beyond). For distances of 20 ft I ran over 50 trials and noticed that the ballon was detected 98% of the time. The confidence score was also very high at >80%. I noticed if I reduced the confidence threshold there were some false positives. I tested the model using other objects too try to break it and the model continued to perform well. I placed objects such as small balls, phones, hoodies and etc. however the model seems very robust and wasn’t affected by the distractions. Overall, I’m very confident in the models ability to work.
Additionally, this week I worked closely with Gaurav to get the kria set up for the vision model. We began by setting up the vitis AI tool, and attempted to load the yolo model on to it. In order to do that, we began by first quantizing the model. In order to do this, we began by exporting the model into a onnx format. Then, wrote a script to quantize the model (from float to int8). Using the quantized model we tried to upload it to the kria following the “quick start” guide. However, we realized that we would have to write our own script that utilized the Kira’s xillinx’s library to analyze the model to ensure it’s compatible with the dpu. Once we wrote the script and ran it on the kria, we kept getting errors that the hardware isn’t compatible. Thus we tried multiple different approaches such as using their onnx library and PyTorch libraries to try to get vitas . After discussing with Varun (FPGA TA), we were recommended to look into using an AI accelerated raspberry pi 5.
We looked into the pros and cons of new hardware and found that is would be much more efficient and would simplify our issues. Working towards fixing this issue in the next week and getting the model working on the new hardware.