15-494/694 Cognitive Robotics: Roboflow Instructions

  1. Log in to your Roboflow account using your CMU Andrew email address. (If necessary, select "log in with Google" and then enter your Andrew address.)

  2. From the navigation bar on the left, select "Projects".

  3. Click on the purple "+ New Project" button.

  4. Give your project a name, e.g., "Lab8", and set the Annotation Group to "dominoes".

  5. Set the Project Type to "Instance Segmentation".

  6. Click on the "Create Public Project" button.

  7. The navigation bar now shows you on the "Upload Data" page. Click on "Select Folder" and navigate to your lab8 folder, select the snapshots folder inside it, then click "Upload".

  8. After the files are uploaded, click on the purple "Save and Continue" button.

  9. Now the navigation bar shows you are on the "Annotate" page.

  10. Choose the "Auto-label Entire Batch" option. You could label the dominoes yourself by drawing polygons, but the built-in labeler is amazingly good at this.

  11. On the "Auto Label" page, enter class name "domino" and visual description "rectangular domino with colored dots". Make sure the labeling mode is set to "polygons", not "masks". Then click the purple "Generate Test Results" button.

  12. Verify that the auto-labeling works on the sample images. If you're satisfied, click on the "Auto Label WIth This Model" button.

  13. Wait for the annotation job to complete. Then, from the Annotate page, click on the "Review" panel.

  14. From the Review page you can approve or reject each annotated image. If everything looks good you can click on the puple "Approve All" button. Otherwise, you can click on individual images to approve or reject them, or manually alter the annotation.

  15. After you've approved the annotated images, click on the purple "Add Approved to Dataset" button.

  16. From the left navigation bar, select the "Dataset" page.

  17. Your dataset needs to be split among Train/Validation/Test sets. Initially all yor images are in the Train set; you need to move at least a few of them into the other two sets. When you hover over an image, a checkbox appears in the upper right corner. Click on the box to select that image. Use this mechanism to select 10 images for your Validation set.

  18. Once at least one image has been selected, an "Actions" pulldown menu appears. When you've selected all your Validation images, click on "Actions" and select "Change Dataset Split", then click on "Set All To Valid". The icon in the bottom left corner of the selected images will change

  19. Repeat this process to select 5 images for your Test set and move them to the "Test" split.

  20. In the left navigation bar, select the "Versions" page.

  21. In the "Preprocessing" step, edit the Resize step to say "Fit to 160x160 with black edges". Then click on the purple Continue button.

  22. In the"Augmentation" step, click the "+ Add Augmentation Step" button and select image level "Rotation". Accept the default value of 15o and click on the purple Apply button.

  23. Click on the purple Continue button to create the first version of your dataset.
  24. In the Create step, select Maximum Version Size of 5x (or 3x if you're concerned about disk space or training time). Set the version name (scroll up to find it) to "rotate-5x". Then click on the purple Create button.

  25. Now on the Versions page you should see the "rotate-3x" version of your dataset. Click on the white "Download Dataset" button and download it to your hard drive.

  26. Set the output format to "YOLO26".

  27. Choose the "Download zip to computer" option and click on the purple Continue button.
  28. Make a lab8/data folder and move the zip file into that folder. Unzip it there.

  29. Now that you've assembled a properly split labeled dataset, the next step is to train your segmenter.


Dave Touretzky