Underweighting empty images?

[asgersvenning]

Here:

flat-bug/src/flat_bug/datasets.py

Lines 214 to 215 in 0f52efe

	self.sample_weights = [image_weight * len(label_i["cls"]) for label_i, image_weight in zip(self.labels, calculate_image_weights(self.im_files))]
	self.__indices = generate_indices(self.sample_weights, target_size=len(self.im_files) * self._oversample_factor)

len(label_i["cls"]) is the number of instances (objects/labels) in the image, so we literally set the resampling weight to zero if the image is empty. This still doesn't mean we completely ignore it, because generate_indices ensures that all images are included at least once per epoch, even if the weight is zero:

flat-bug/src/flat_bug/datasets.py

Lines 83 to 84 in 0f52efe

	for i, w in enumerate(weights):
	indices.extend([i] * int(max(round(w), 1)))

but still, perhaps we should change to:

self.sample_weights = [image_weight * (1 + len(label_i["cls"])) for label_i, image_weight in zip(self.labels, calculate_image_weights(self.im_files))]

Should try out when we train V2.

[qgeissmann]

Just a clarification, during data preparation, we do not /should keep images with no instances, do we?

[asgersvenning]

I agree that if we want to remove empty images, that should be done during data preparation, but I would think that they don't have zero value, even if they aren't as helpful as images with instances.

Probably better to balance this manually when choosing which images to annotate?

But I guess what the optimal balance will be an open question for now.

[qgeissmann]

What I mean is that we already remove empty images during preparation, so, no empty image should be seen by YOLO, the least I recall (fb_prepare_data.py). On the one hand, this is not ideal as BG-only images have information. On the other hand, it makes it easier to handle partially annotated datasets with he convention that images with no annotations are not yet annotated...