Add SODA-A dataset

[nilsleh]

This PR adds the SODA-A dataset. Dataset rehosted on HF.

Dataset features:

* 2513 images
* 872,069 annotations with oriented bounding boxes
* 9 classes

Dataset format:

* Images are three channel .jpg files.
* Annotations are in json files

TODOS:

• some annotations have more than just 8 coordinates, for example file Annotations/train/01874.json:

{
        {
            "poly": [
                1248.874059994201,
                2785.0,
                1267.2565270181412,
                2785.0,
                1290.0001220703125,
                2783.999755859375,
                1289.2877197265625,
                2767.801025390625,
                1248.1971435546875,
                2769.608154296875
            ],
            "area": 659.4863047840048,
            "category_id": 7,
            "image_id": 1874,
            "id": 43
        },

Example plot:

[nilsleh]

@shaunyuan22 Thank you for this nice dataset and all the work. We aim to make the dataset more easily usable in torchgeo, and would appreciate it if you have any comments, corrections etc.

[nilsleh]

@adamjstewart what should we do for oriented bounding boxes, Kornia only has boxes_xyxy and boxes_xywh.

[ashnair1]

Kornia box keys are bbox_xyxy and bbox_xywh

[nilsleh]

Open question how to deal with the polygons into a common oriented bounding box schema.

[adamjstewart]

Asked about oriented bboxes on Slack. We may need to add support to Kornia for this ourselves. Until Kornia supports it natively, I guess it doesn't matter what the format looks like. But let's use the same key names that Kornia uses.

[adamjstewart]

Suggested change

	`SODAA`_,OD,Aerial,"CC-BY-SA","2513",9,"varying","RGB"
	`SODA`_,OD,Aerial,"CC-BY-SA","2513",9,"varying","RGB"

I see CC-BY-SA here, but I wish we knew which CC-BY-SA version. I usually use a size range instead of "varying".

[adamjstewart]

@shaunyuan22 do you know?

P.S. We are adding TorchGeo data loaders for your excellent SODA-A dataset, hopefully this makes it even easier for people to use and cite your paper!

[adamjstewart]

Should we mention that pyarrow is required? Is this a requirement from the dataset authors or from you?

[nilsleh]

From me

[adamjstewart]

I'm hesitant to add extra dependencies unless they are absolutely required. Does it add a significant speed boost?

[nilsleh]

Alright, Ill change it to csv, parquet is more performant so today's standard, but that will remove the dependency.

[shaunyuan22]

1. For the license, our SODA-A dataset is under CC BY-NC 4.0 (see here), which is stated in the Appendix of our paper;
2. For the preprocessing of polygons to oriented boxes: we use OpenCV to perform such convertion:

# ann: an 8-point polygon
poly = np.array(ann['poly']).astype(np.float32).reshape(-1, 2)
# obbBox coems as standard 5-parameter representation in oriented object detection: [x, y, w, h, a]
# where `x` and `y` refer to the center point of the oriented box, while `w` and `h` formulate the 
# width and height of this instance, and `a` stands for the orientation angle
obbBox = cv2.minAreaRect(poly)
# we can also capture its 8-point polygon format of the oriented box
obbPoly = cv2.boxPoints(obbBox)

Big thanks for your attention to our work! Let me know if you need anything.

[adamjstewart]

@shaunyuan22 the Appendix says CC-BY-SA but links to CC-BY-NC-4.0. One of these is a typo, but I'm not sure which.

P.S. We have a similar question about DIOR if you or anyone else from your lab knows: #2572

[shaunyuan22]

Sorry for the typo and unclear description, our datasets (including DIOR and SODA) are under CC BY-NC 4.0 license.