Post processing mask refinment

[qgeissmann]

ATM, when we want to annotate more images, we use flatbug to generate candidate annotations and then correct them, so we want high-quality masks (there are other contexts in which we want high-quality masks).
What happens now, is we have poor mask resolution for small objects.
For instance, the grid here is 500 px.

In my use case, either I spend a lot of time fixing masks by hand or axxept low quality, but then poison the dataset...

For cases where we need high quality mask, it would be good to have a post-processing option.

For another project, I managed to get better masks by

1. finding instances
2. cropping instances as a square with a + a margin as a square
3. scalling the square to 1024 px
4. running the same yolo model on this image
5. if finding a new mask for this instance (e.g. based on iou), keep the next mask (else fall back on old one)

Possibly we can work on subtleties such as masking non-focus detections

I propose, for now, to implement this as a separate tool, we can merge later if it makes sense...

[qgeissmann]

promissing, but not quite there yet, blue is new proposal (8e23f7f):

[qgeissmann]

it works beautifully for some datasets such as "blair2020", which were so far prospective as too long to annotate well:

[qgeissmann]

it does not always work so well on small instance:

I think it is because flatbug does not see such case during training as we pass an image at native scale or zoomed out, not in.
One way to improve this either in flatbug, or in a separate model would be to take high-quality masks on large instances and augment them by downscaling and reupscaling the images.

Possibilities are:

1. implement as possible augmentation in current flatbug (I have ideas, maybe)
2. trained a derived model that does that (fine tune yolo weights on such scenarios) and have a refiner-only model. Possibly, such a model could be smaller/faster yolo?