TFLite model validate currently requires changing the script and running in debugger mode to obtain the output equation. Change the program so that it runs as the following when running like ../../scripts/tflite_model_validate --image=<path_to_image> --ptiq_path=<path to tflite w/ PTIQ> --no_ptiq_path=<path to tflite w/o PTIQ>

1. Load the image
2. Load PTIQ and non-PTIQ model
3. List model outputs. Ask the user which output is the boxes, classes, scores, and num_boxes.
4. Run inference on the non_PTIQ model. Draw bounding boxes. Show boxes to user and ask if ok. Save bounding box data.
5. Run inference on the PTIQ. Use linear regression to create the equation. Show bounding boxes to user. Ask if ok.
6. If user answered "ok", export the equation to the user.

Currently we are using OpenCV from apt repo. Ideally, we should be building OpenCV from source so that we can build with -O3 optimizations and control which version we use.

Create documentation for a non-developer to be able to use OrbitVision inference

Create README describing model use cases

Currently we are using libedgetpu from apt repo. Ideally, we should be building OpenCV from source so that we can build with -O3 optimizations and control which version we use.

Model Name: FRC_2022

Number of Labels: 2

Label Names: red_ball, blue_ball

Base Model: ssd_mobilenet_v2_320x320_coco17_tpu-8

Improve the documentation for exporting to TFLite by including specific commands for how to export

Flash limelight with, and test OrbitVision disk image. The LL is a Pi Compute and should work with our code

convert_to_tflite should accept the exported model path and model name, and create both a PTIQ and non_PTIQ model in exported_models/. Script should be run by doing something along the lines of:

../../scripts/convert_to_tflite.py --export_model_path=<path to exported model> --model_name=<model name>

Currently an RPi is needed in order to build and develop the inferencing application. This is not ideal, as it is slow to build and develop. Ideally, we can have a VM running Raspberry Pi OS, and use that to build and develop. Research a way to do this.

Write values (bounding box coordinates, label, angle to target, etc) to NetworkTables so RoboRIO can access inference results.

Create a webserver so a computer connected to the robot's network can view the camera feed, both with and without bounding boxes

Add more information on the initial README file

Script should

• Change the CPU governor to "performance"
• Verify that the TPU is connected
• Verify that the camera is connected
• Launch OrbitVision