Tile Images and Annotations for large images

[davodogster]

A common problem for researchers is having large images with lots of small objects. Is there a function in datumaro to tile images (overlapping option possible, min_visibility option possible) into specified dimensions and the corresponding masks, bounding boxes, and polygons also get tiled, with images renamed with their original coordinates. If not, then it would be great to have this functionality.

Cheers, Sam

[zhiltsov-max]

Hi, there is no such function yet, but it is an interesting feature request. If you have an opportunity, you can try to implement it yourself - it would be a Transform, and it does not look like something very complex to implement. How do you think images should be tiled - is it some even split, or a cut of a specified size?

[davodogster]

Hi, there is no such function yet, but it is an interesting feature request. If you have an opportunity, you can try to implement it yourself - it would be a Transform, and it does not look like something very complex to implement. How do you think images should be tiled - is it some even split, or a cut of a specified size?

@zhiltsov-max Hi, I think that the user should have control over how they want the tiling/slicing to be done, so they can either chose (n_rows, n_columns) or (tile height, width), with an overlap size of 0 to 500+. Once the tiling of images and all annotations are performed then the function outputs a newly annotated dataset.

[nmanovic]

@davodogster , @zhiltsov-max , you can find a nice discussion of the problem here: cvat-ai/cvat#531. I think that Datumaro should not tile an image (at least as next step). At the same time it should be able to work with such images (e.g. extract crops with objects, import/export datasets with big images, etc). Let's keep the issue in backlog.

[davodogster]

@nmanovic cvat#531 No my problem isn't loading large images into cvat. The problem is that it would be nice to tile the images and corresponding annotations after they have been annotated to increase the size of the dataset to split into train/val/test, and it also means that image resolution can be preserved so we can train on the full resolution after tiling and don't need to downsample the resolution.

[collinrpower]

@davodogster I have run into the same problem. Have you been able to figure it out?

[michaelSkaro]

I also have the same problem. Is there an implementation?

[zhiltsov-max]

@michaelSkaro, not yet. Probably, we need better understanding about how this feature should be designed. If possible, could you tell us more about your use case and possible solutions you see?

[mmoollllee]

I'd love to see a functionality like this in CVAT: https://github.com/nagi1995/yolo-tiling/blob/main/tile_yolo.py

I have a dataset with highres images with few small objects. I want to tile every image to get rid of the image parts without bounding boxes but keep the original files.