I am having problem in start training Object Detection model on Google Cloud Vision. I have uploaded 342 images of 4 classes, and labelled 102 of them. There are 38,24,20,20 labelled images for each class some of which have multiple bounding boxes. In the Label Stats pane, it says each label should have at least 10 bounding boxes, and at least 8, 1, 1 bounding boxes each. It also writes "Your dataset will be automatically split into Train, Validation, and Test sets." But all my labels stay in Train column and val and test columns are zero. Is there any other way to split them manually, or am I missing a point?
Can't start training in Google Cloud Vision Object Detection, Train Test Validation not splitting
Viewed 1805
3 Answers
If you see this message:
Some of your labels (e.g., "") do not have enough bounding boxes assigned to your Validation / Test sets. Import another CSV file and assign those images to those sets.
when you click on the Train tab on Google Cloud Platform, your problem is likely that you uploaded images and only labeled some of them. When you first upload your images, Google assigns them to a train, test and validation sets - before they have been labeled. So this message is misleading:
Your dataset will be automatically splits them into Train, Validation, and Test sets.
What is happening is that all of your images have been split, including those images that you haven't labeled yet. The test and validation images are at the end of the list, and if you labeled in order, then those are not yet labeled, and this error is triggered.
If you label all of your images, or only upload images that you plan to label it should work. You could try removing the unlabeled images as well.
You can create a CSV file containing the image URI and labels including to which set (e.g. TRAIN, VALIDATION or TEST) does the image belongs to. Reference
If you think you have enough images and have done the bounding boxes manually via the browser you can use the export button to create the CSV file and go in and modify some of the training images to validation
Related