Training using nongeoreferenced Airbus Ship detection data

[julled]

Hi, i just found your very nice framework. Its nice that its specifically tailored for satellite images.

I am wondering how you recommend the usage of the framework for the usage on non-geotiff / nongeoreferenced / png data like the Airbus Ship detection data. We want to use the data in a object detection task.
We converted the segmentation labels to a coco json label file with one rectangle per boat instance.

I already found out that its possible to load the pngs using a rasteriosource.

What is now unclear to me is how to feed in the coco boat rectangle labels for usage in a ObjectDetection pipeline. One option is to convert the coco jsons to geojsons. Is it possible to use geojson files with labels in pixel coordinates in combination with the rasteriosource?

Or do you recommend any other format ?

Thank you for your help!

[AdeelH]

Hi, I suggest using the master branch.

You do not need to use RasterioSource and you do not need to convert annotations to GeoJSON; RV can handle pre-chipped data. You will need to create an ObjectDetectionImageDataset. This requires your annotations to be a JSON file in the COCO format.

Once you have the dataset, you can proceed with training as normal. See this tutorial on training. Note where train_ds and val_ds are being defined. You will need to change these to be ObjectDetectionImageDatasets. You will also need to replace the semantic segmentation-specific classes with equivalent object detection ones.

Also see this discussion.

Let us know if you run into any trouble.

[julled]

Thanks for the detailed answer, this helps alot already!

Our usecase would be now to use the Airbus dataset as a base source for data and after training do a finetune on a smaller custom dataset. Is finetuning possible using rastervision?

[lewfish]

Yes, you can finetune. See https://docs.rastervision.io/en/latest/framework/misc.html#transfer-learning-using-models-trained-by-rv

[AdeelH]

@julled

While preparing this notebook, I realized that although the Airbus aircraft detection dataset has non-GeoTIFF images, those images are still large enough (2K+ by 2K+ pixels) to require chipping.

If this is also true for the Airbus dataset you are using, then, yes, you should be using RasterioSource and not the ImageDataset as I said in my earlier comment.

I can try to provide more guidance if you share some more details about your dataset.