facebookresearch
SAT-493M pre-trained segmentation head weights are not provided
When using the SAT-493M pre-trained backbone_weights, I found that the repository does not seem to provide the segmentation head weights pre-trained based on this dataset. Thanks!
Which pretrained head are you interested in? Eg there is not a SAT493M on coco.
The geo bench heads are quick to train with eg TerraTorch and the licensing for iSAID, DIOR does not allow for releasing models trained on the imagery as they have academic only licenses.
The updated canopy height model on the expanded SatLidar dataset will be released at a later date.
I am happy to help provide suggestions on how to recreate any of the models, though!
No all the geospatial benchmarks use DPT or UperNet as noted in the appendix!
I really want to know whether the SAT-493M model supports 512×512 remote sensing image input without being resized to 224×224 by the image processor. Is it possible to directly modify the preprocessor_config.json of the SAT-493M model to 512×512? Thank you very much for your answer.
Hi @kb077 -- because DINOv3 uses rope embeddings, it can take in images of any resolution -- indeed the SAT493 model had high resolution fine tuning at 512x512 resolution. You can definitely adjust it to take in 512x512 images while keeping the backbone frozen. DINOv3 is SOTA on iSAID, which trains on 896x896 patches, with a frozen backbone.
Has anyone worked on multi class segmentation on satellite images ( high resolution & Low resolution), currently im working on Urban Informal (Slum) detection, as it will be pixel wise classification how can i utlize DINOv3 for my downstream task
