Mail communication between Prof Michael and myself.
Things I’ve done since the last meeting:
More search on existing literatures of automatic scoring systems. – They are all based on image processing.
Started learning TensorFlow.
Content we talked about during today’s meeting:
Size of the training images: do experiments, start to reduce resolution, and see how long it takes until it starts to lose precision.
Label is separate from the training images, we are not actually drawing marks or bounding boxes to the image, it’s a label thing like the TensorFlow fashion examples. Think thoroughly about how to make use of AR kit to collect data, or even can AR element solve all our problem? – what is the goal of this project.
For the annotations, only have the x on the tip of the arrows, and maybe add the ring number. Because not every time you can see the entire arrow in the photo.
A common approach for most neural nets is to have a fixed input size image. So do some classic image processing first (cropping, deskewing the images) before feed them to the network. When using the networks we train, we are also going to do whatever basic computer vision you do on the training data. e.g., if you convert images to B&W when doing the training, you should also make them B&W when using the camera in the field.
Search for potentially existed image set of archery target. – They are probably not annotated.
What I plan to do next:
Keep learning and search for deep learning related material.
Search for existing archery data set.
Think through the project – project planning, think of the goal of the project, if some existing tool like AR thingy can solve most of the problem.
Draw a system architecture diagram, components of the project. Send the draft for PDD before the next meeting, which is Tuesday, two weeks from now.
Gantt charts can be useful for time planning.