As I get better at training FOMO. One of the tricks for me has been to increase the amount of backgrounds behind the bounding boxes. It occurs to me that if I set bounding boxes to objects on a translucent background image, it would be very useful if edgeimpusle filled in the background with multiple different colors or random backgrounds. Presently I don’t think edgeimpulse allows images with translucent background.
Actually changing the background can be done fairly easily manually by.
- removing all items from the test set to the training set
- load into the test set an image of objects with a continuous background color like white
- set your bounding boxes
- make sure this is the only image in the test set
- export your data, backup the testing set folder.
- in the testing set folder, (has the image and bounding box JSON data) change the background color, save the image (don’t change file names).
- upload the image and bounding box info to the test set
- check it arrived and repeat
- when all images with new backgrounds have been uploaded move all images to training set
- rebalance dataset.
Good find Something that I like to do is write a Google Colab script that will perform data augmentation automatically so that I can just send it a few samples and it will mix in different backgrounds (I did this for keyword spotting, but I have to imagine it’s possible for images). You can then use the EI uploader to automatically send everything to your Studio project.