I would like to know a bit more about what is happening in terms of prompt to segmentation mask when using PCS with Seg Preview (RF-DETR) on Roboflow Playground.
I’ve taken an image that I would typically have to run custom training or refinement on if using other models, and for which classes do NOT exist in the COCO class list, given the class as a prompt and it’s got nearly perfect segmentation.
I’ve been trying to replicate this in code within my own workbooks using the preview model from the very anaemic examples and code provided, and for the life of me cannot work out what is being used for prompt to mask. The performance in playground is excellent, but I’ll likely need to iterate over hundreds of thousands of images in a a highly customised workflow, so would love to be able to do this locally.