Add Spacenet 1: Building Detection v1 #129
Conversation
|
|
There was a problem hiding this comment.
Thanks for the PR!
For the tests, can you replace the example images with fake images? See https://github.com/microsoft/torchgeo/tree/main/tests/data#raster-data for details on how to create them. We can't redistribute dataset files without worrying about licensing, and larger files mean slower downloads.
ashnair1
force-pushed
the
spacenet
branch
2 times, most recently
from
704c291 to
f07741e
Compare
I've created the dataset using the specified instructions though I had to hack together a script to do so. Incidentally, I would like to discuss the test dataset generation process but I'll raise a separate issue for that. |
|
Currently the three remaining issues are those concerning:
I'm looking into Regarding issue 3 (which is also related to the creation of a Spacenet meta class ), this is dependent on how much we would like to control what data Other spacenet datasets also contain multiple imagery. For example, spacenet 2, 4 and 5 provide MS (Multispectral), PAN (Panchromatic), PS-MS (Pansharpened Multispectral) and PS-RGB (Pansharpened RGB , though this can be created from PS-MS). Spacenet 6 provides SAR and EO imagery. Spacenet 7 is a time series dataset. Some spacenet datasets are more similar than others but the quality of having more than one input image is pretty consistent. I'm not too familiar with Spacenet 6 and 7 but for the rest, people generally tend to use the PS-RGB when available since that's most compatible with pretrained backbones. So we can either provide a subset of the data i.e. the one image that is commonly used in all of the winning solutions or all of it. The former will make implementation easier (probably make a single spacenet file for all the datasets) but the latter is a more true representation of the full dataset. I'm open to either option. |
|
We discussed this in a meeting today and it seems like the easiest way to move forward would be to keep this as a VisionDataset with an integer |
* Create single spacenet.py for all spacenet datasets * Create single spacenet directory for all spacenet test data * Create single test_spacenet.py for testing all spacenet datasets
|
Sure, makes sense. I've made the necessary changes. |
|
Since Spacenet is currently implemented as a |
|
Just a heads up -- I tested this dataset and found that the "rgb" images come in three different shapes: {(3, 406, 438), (3, 406, 439), (3, 407, 439)} while the "8band" images are 1/4 the size. It looks like this is actually expected -- https://medium.com/the-downlinq/getting-started-with-spacenet-data-827fd2ec9f53 -- but it means that you can't immediately use it with a dataloader (and will instead need to do some resizing in a transform). Do we want to address that anywhere or let people figure it out? @adamjstewart is the only thing missing here the "Spacenet" --> "SpaceNet" find and replace? |
|
Since it's only a 1 pixel difference I think we should handle it ourselves. Yes, the only thing remaining is the rename. |
|
Thanks for the review. Rename complete. |
* Add Spacenet 1 * Add test data * Style fixes * Convert Spacenet1 to VisionDataset * Add option for selecting imagery * Consolidate spacenet * Create single spacenet.py for all spacenet datasets * Create single spacenet directory for all spacenet test data * Create single test_spacenet.py for testing all spacenet datasets * Add copyright * Reorder Spacenet in docs * Test both rgb & 8band * Rename Spacenet -> SpaceNet
* Add Spacenet 1 * Add test data * Style fixes * Convert Spacenet1 to VisionDataset * Add option for selecting imagery * Consolidate spacenet * Create single spacenet.py for all spacenet datasets * Create single spacenet directory for all spacenet test data * Create single test_spacenet.py for testing all spacenet datasets * Add copyright * Reorder Spacenet in docs * Test both rgb & 8band * Rename Spacenet -> SpaceNet
torchgeo.datasets.Spacenet1from SpaceNet: A Remote Sensing Dataset and Challenge Series. Dataset homepage - here