I'm deciding whether or not to use Prodigy. Does Prodigy have a Re-ID workflow? That would involve showing an anchor image, then generate a 'selection set' of images that to be ranked as 'same' or 'different'. This could be semi-automated by generating only the selection set that produces the optimal model performance improvements, instead of having to do every combination of people comparisons.
Hi! The workflow sounds like a pretty classic use case for Prodigy, yes. You'd just have to implement the stream and logic that you want. Here are some relevant docs pages with more details:
- Custom recipes for scripting annotation workflows: https://prodi.gy/docs/custom-recipes
- Computer vision guide with examples of different UIs for presenting images and how to put it all together in a recipe script: https://prodi.gy/docs/computer-vision
For instance, your stream could load the anchor images and options and display them in a multiple choice interface and/or pre-select options, or decide which options to present based on the image. You could also train an annotation model in the loop only for making the pre-selection and then update it with the answers. You might have to experiment a bit to find the best compromise of preprocessing, batch sizes etc. if you want to use an image model in the loop and actually update it – your model may be too slow to do everything at runtime, and you might have to adjust the implementation to make it sensitive enough to small updates, so you see results quicker.