Has segment-geospatial finetuned SAM on a remote sensing dataset?

[adamschneider42]

I am just wondering if the SAM has been finetuned on a remote sensing dataset, or the code has just been modified to handle geotiffs and geospatial data?

[giswqs]

Not yet, but contributions are welcome. See #35

[amalshehan]

For now, we can only fine-tune what SAM allows. See the parameters below. It might take sometime before we can support other model fine-tuning. Contributions are always welcome.

https://github.com/opengeos/segment-geospatial/blob/main/samgeo/samgeo.py#L43

  points_per_side: Optional[int] = 32,
  points_per_batch: int = 64,
  pred_iou_thresh: float = 0.88,
  stability_score_thresh: float = 0.95,
  stability_score_offset: float = 1.0,
  box_nms_thresh: float = 0.7,
  crop_n_layers: int = 0,
  crop_nms_thresh: float = 0.7,
  crop_overlap_ratio: float = 512 / 1500,
  crop_n_points_downscale_factor: int = 1,
  point_grids: Optional[List[np.ndarray]] = None,
  min_mask_region_area: int = 0,
  output_mode: str = "binary_mask",

Any examples of tuning with these parameters? samgeo seems to be unable to pass the parameters to sam. Has anyone else had a similar issue ? It is throwing a unexpected keyword argument error!

[giswqs]

So is this a samgeo issue or the original SAM issue? samgeo already has sam_kwargs in the SamPredictor constructor.

if automatic:
    # Segment the entire image using the automatic mask generator
    self.mask_generator = SamAutomaticMaskGenerator(self.sam, **sam_kwargs)
else:
    # Segment selected objects using input prompts
    self.predictor = SamPredictor(self.sam, **sam_kwargs)

[amalshehan]

The issue is on both sides. In the original SAM SamPredictor does not take kwargs so better to remove the **sam_kwargs from the call to it in samgeo.

[giswqs]

It appears to be the issue of the original SAM SamPredictor. I will keep sam_kwargs in the samgeo function in case SAM supports it in the future.

[amalshehan]

@giswqs Ok thanks for looking into this. If possible better to update the geosam docs. As you suggest good to keep it since it is vital to have that option when using it on diverse types of data. Is there a workaround to this? I recall ArcPro deployment allowing parameter tuning!

[amalshehan]

@giswqs any plans to improve samgeo based on sam-hq?