No announcement yet.

Online Collabration for Dataset Generation - Object Detection

This topic is closed.
  • Filter
  • Time
  • Show
Clear All
new posts

  • Online Collabration for Dataset Generation - Object Detection

    Hello everyone!
    This year's objects are a bit non-conventional, and the better way to deal with it is using Deep Learning for object detection. There are approx ten classes of objects this RoboSub, and such a huge dataset creation is resource consuming. I got the idea from last year's forums, how we could collaborate to build this dataset, with different underwater camera qualities, pool lighting conditions and image prints.
    If some of the teams are up for it, we can definitely set up a GDrive/Dropbox link which would enable us to contribute to this. (if this doesn't violate any rule, i don't really think it would)

  • #2
    Wait.... all of my previous objects were ...conventional...?

    Excellent idea.


    • #3
      Triton Robosub would love to contribute!


      • #4
        Has there been any new information regarding the potential shared dataset?


        • #5
          Hey everyone!

          OSU's UWRT has decided to start off this effort. Below is a link to a google drive folder with all of our footage so far. Anyone can modify the folder as they see fit and use our data. All of these videos should have the props without them leaving frame. We only have one video of the path marker from last year.

          Feel free to add any footage and let us know if you have any questions or requests related to training or videos!

          With love,


          • #6


            • #7
              Hi guys, a bit of advice for data collection.
              So to get your deep learning algorithm to have a greater detection accuracy vvv
              You'd want to collect data that varies in rotation and translation as well. You want there to be some variance.
              Perfom data augmentation( like blurring, change in contrast, rotation) and shuffle the data up.

              If you're noticing the loss is low and you still have a lot of false positives, you can perform hard negative mining to reduce the false positives.

              Long story short, for hard negative mining: draw a bounding box around the false positive detections, label them as 'not_object', feed this new knowledge into the training process.

              - UCR Robosub


              • #8
                OSU_UWRT Thanks for kicking it off. We would also like to share the videos that we have collected from testing our submarine. I will add more videos to this drive folder as we continue to test our vehicle.


                - Tartan AUV


                • #9
                  rishj09 Thanks for those videos. I'm sure you can tell from our data but we lost our Aswang prop. We also need footage of the garlic drop if anyone has that. We are uploading more videos we have taken since the original upload.



                  • #10
                    Hello. Here is some images, bins mostly. Hope it will help you.



                    • #11
                      Just wanted to let everyone know we uploaded our training footage we recorded today to the google drive link (! We talked to several teams already and encourage everyone to upload their footage!

                      Good luck everyone!
                      Last edited by OSU_UWRT; 07-30-2019, 08:02 PM.


                      • #12
                        Hi everyone! Great idea to share data around. We'll be updating this folder with footage as we take it throughout the week. Hope it helps!



                        • #13
                          ARVP from the University of Alberta is sharing some footage here:
                          So far, I think the footage is mostly just buoys, torpedo, and gate. Hope this helps!
                          Last edited by NinjaPerson24119; 08-01-2019, 12:43 AM.