[TASK] Train Yolo Model For Valve Detection

[jorgenfj]

Description of task

Training

We need a robust way to detect valves during TACC. This issue is solely dedicated to training the model. Deep learning models require lots of annotated data in over to work effectively. Therefore we should start this process as soon as possible.

Generating Data

We have a 3d print of the valve at the office which can be used to generate a dataset. It might also be wise to look into the creation of a synthetic dataset in blender with automatic labeling.

Remember that the dataset should reflect real world conditions. There is a rosbag on teams from the pipeline following mission run from this summer. By looking at that one could get a feel for how the operating conditions are underwater (murky water). Sadly there are no recording of the valve manipulation task.

Suggested Workflow

• Talk to @Senja20 about how to start this task. He has experience training a yolo model with Roboflow.
• Roboflow is a good tool for labeling data. Should look into that. How to Train YOLOv8 Object Detection on a Custom Dataset
• Look into the possibility of creating a Synthetic Dataset in Blender

Specifications

• The valve is static, so accuracy should be somewhat prioritized over speed.
• Should be able to input an image and generate bounding boxes for detections

Contacts

@Senja20 @jorgenfj

Code Quality

• [ ] Every function in header files are documented (inputs/returns/exceptions)
• [ ] The project has automated tests that cover MOST of the functions and branches in functions (pytest/gtest)
• [ ] The code is documented on the wiki (provide link)

[jorgenfj]

https://github.com/atenpas/gpd https://github.com/lianghongzhuo/PointNetGPD https://jeasinema.github.io/file/pointnetgpd_icra19.pdf

[jorgenfj]

https://pcl.readthedocs.io/projects/tutorials/en/latest/correspondence_grouping.html

[jorgenfj]

https://pcl.readthedocs.io/projects/tutorials/en/latest/global_hypothesis_verification.html#global-hypothesis-verification