Convert D-Fine to CoreML

Hello,

I am currently working on an iOS application that requires object detection functionality, and I’m interested in using the D-Fine model for this purpose. I would like to ask if anyone has experience or knows of a way to transfer the D-Fine model into the CoreML format to be used on iOS devices.

Thank you in advance!

[olokobayusuf]

@ProfessorHT I can probably help here. If you can write the DFINE inference as a Python function that takes in an image and returns detections, then we can compile it to run on iOS (including converting the model to CoreML) using Muna.

You can also try out RF-DETR in your iOS app. Try the demo here then see the "API" tab on how to import the model into your iOS (Swift) app.

Feel free to follow up on our Slack. We can also help you write a Python function that runs DFINE so that we can compile it for iOS.

[isaacrob-roboflow]

@olokobayusuf RF-DETR doesn't need NMS, but your code applies NMS

[codeperfectplus]

I converted the model and tested it on a Mac Mini M4. With the original DFinex model, I was getting around 7 FPS, while the converted version achieves 16–24 FPS depending on the input size (480 or 640). The precision, recall, and F1 score remain identical to the original model. You can check the code in my fork of the repository.

[ArgoHA]

you can also try D-FINE-seg, it has make export that includes CoreML out of the box (int8 version too)