Transforms v2 failing for real Coco Dataset

data
1

I’ve recently tried to run the Transforms v2 tutorial on end-to-end object detection.
When working on the real Coco 2017 train dataset, v2.SanitizeBoundingBoxes always fails due to some coco samples not containing any annotations.

Is it really not possible to use Transforms v2 with CocoDetection out-of-the-box, like described in the tutorial or am I missing something obvious?

Thanks in advance for your help,
Benjamin

2

So it seems like there exists a utility function for coco to remove the unlabeled datapoints here. Unfortunately, this does not play together with the datasets.wrap_dataset_for_transforms_v2, but can be used in an own dataset class. I’m still perplexed as to why this is not really mentioned anywhere.