You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
STAC is being used quite a bit for other datasets that might be useful to OCF and to general datapipes. The Clay EO foundation model runs off STAC for getting the satellite imagery, and more and more STAC catalogs are being added to Source Collective, Planetary Computer, etc. which might be helpful for OCF.
Detailed Description
There is https://github.com/Clay-foundation/stacchip that generates image chips from STAC catalogs. They also do a preprocessing to batches first approach, rather than streaming it in from what I can tell, but the flexibility in their batches could be something that might be worth trying? Its similar to NumpyBatch, but more flexible.
Context
Clay is a foundation model being built to be flexible with a variety of different inputs, currently mostly imagery, but also topographic maps, and the imagery is both satellite and aerial. The model can handle if any arbitrary number of bands are available or not available, as well as different resolutions, using some cool tricks from https://github.com/zhu-xlab/DOFA and the flexibility of the transformer. Their data loading has many of the same problems that datapipes is supposed to handle, although mostly using STAC and GeoTIFF vs Zarr, but could be some interesting things to look at.
Possible Implementation
Rather than NumpyBatch with Enum's that we have to track, you could do just a normal dictionary, like Clay does.
The text was updated successfully, but these errors were encountered:
STAC is being used quite a bit for other datasets that might be useful to OCF and to general datapipes. The Clay EO foundation model runs off STAC for getting the satellite imagery, and more and more STAC catalogs are being added to Source Collective, Planetary Computer, etc. which might be helpful for OCF.
Detailed Description
There is https://github.com/Clay-foundation/stacchip that generates image chips from STAC catalogs. They also do a preprocessing to batches first approach, rather than streaming it in from what I can tell, but the flexibility in their batches could be something that might be worth trying? Its similar to NumpyBatch, but more flexible.
Context
Clay is a foundation model being built to be flexible with a variety of different inputs, currently mostly imagery, but also topographic maps, and the imagery is both satellite and aerial. The model can handle if any arbitrary number of bands are available or not available, as well as different resolutions, using some cool tricks from https://github.com/zhu-xlab/DOFA and the flexibility of the transformer. Their data loading has many of the same problems that datapipes is supposed to handle, although mostly using STAC and GeoTIFF vs Zarr, but could be some interesting things to look at.
Possible Implementation
Rather than NumpyBatch with Enum's that we have to track, you could do just a normal dictionary, like Clay does.
The text was updated successfully, but these errors were encountered: