Multiband tiff more than 4 bands in rastervision

[ShikhaGupta99]

Hi,

I am currently working on an object detection project using high-resolution TIFF imagery with 6 spectral bands along with corresponding GeoJSON labels for training. I went through the documentation and tutorials, which are indeed helpful, but I could not find detailed guidance on:

1.Using multispectral TIFFs (>3 bands) for object detection.
2.Adapting or selecting different detection models or backbones within Raster Vision.
3. I would be extremely grateful if you could share another document specifying more details and kindly guide me on:
4. How to set up the pipeline to read all 6 bands from TIFF imagery for training detection models.
5.Whether Raster Vision supports switching to different detection architectures (e.g. YOLO, YOLT) or if Faster R-CNN is the default option currently.
6.Any example configurations or best practices when working with multispectral inputs for detection. Also if i can use rotated detection.

I am genuinely fascinated by Raster Vision’s design and its integration capabilities with geospatial data formats. It would be really helpful, if you could share github repo on the projects which have used the framework for object detection bands more than 3. For the guide already present in the document was not that useful for me.

Looking forward to your response.

[AdeelH]

Hi, thanks for the kind words!

• Multispectral imagery should work out of the box. See e.g. the multiband option in this example.
  • This is another example of reading a multiband image into an OD dataset (if you're using Raster Vision as a library instead of the full pipeline).
• Raster Vision uses FasterRCNN by default. Using a different model is possible but trickier. See this example for more details.

Also if i can use rotated detection.

There is no built-in support for that.