Vision Processing Pipeline

[alecGraves]

• Decide on pipeline setup, network architectures, datasets

• Neural networks process and run on image data

  • Make a detection library in tensorflow
  • make class to easily run model
  • make a ros wrapper for the library

• Collect new data

  • Set up qualification pool obstacles
  • Manual control of the sub / low-def gopro footage

[alecGraves]

@eneleski Step 2: python-based data lebeling tool,

• draw boxes over different categories of images

[alecGraves]

data labeling tool moved to #13

[hjevans]

Github documentation for TensorFlow object detection API models, etc:
https://github.com/tensorflow/models/tree/master/research/object_detection

G-RMI powerpoint w/ info
http://presentations.cocodataset.org/Places17-GMRI.pdf

Tiny YOLO (general YOLO webpage):
https://pjreddie.com/darknet/yolo/

[alecGraves]

just an idea: use batching to more efficiently process data from multiple cameras at the same time
And maybe use a recurrent network to help the results stay continuous over time... If we do not do that, the results will be super jumpy, using tinyyolo.

Also, we should try using a custom tinyyolo architecture based on mobilenet to improve accuracy. Or a custom NAS- based model. Basically swap base models, train on mscoco, and fine-tune for our application. We will need to use tensorflow to maintain compatibility with the movidius.

[alecGraves]

TODO: Get retinanet to work, try swapping backend, etc.

[alecGraves]

We are transitioning to the movidius as the main processor for the underwater competition.

@cmcdermitt is working on making an object-detection neural network work, @eneleski and @zcesar1 are setting up the movidius sdk and getting ready to work on movidius ros stuff.