Is there a onnx object detection model available?

[kumardesappan]

Hi, Is there a pre-trained onnx object detection model available for reference (SSD, Faster-RCNN, RetinaNet or Mask RCNN)?

Thanks and Regards, Kumar.D

[houseroad]

We are working on this. To import these models to ONNX model zoo, we may need to introduce some new ops into ONNX spec.

[kumardesappan]

@houseroad Thanks for the update. Do you have any tentative timeline for this?

[houseroad]

Our plan is to add these models in 2~3 weeks.

[prasanthpul]

@houseroad Can you share list of models you plan to add?

[houseroad]

Our plan is to convert the models from https://github.com/caffe2/models.

bvlc_googlenet, bvlc_reference_caffenet, bvlc_reference_rcnn_ilsvrc13, finetune_flickr_style will be converted.

And many models in sytle_transfer and detectron (Faster-RCNN included) folders are also in our plan.

[0wu]

Is there a ticket where we can peek the new Ops for detection?

[houseroad]

Will send the PR including RoiAlign, BBoxTransform, BoxWithNMSLimit, and GenerateProposals which are need by Faster-RCNN soon.

[reneschulte]

Any update on this like an ETA? :)

[0wu]

@houseroad any updates? or possibility to contribute code for this?

[houseroad]

Sorry, the task got delayed,, will send PR soon to the main repo to add these ops.

[bestlin]

@houseroad do you have some plans for the detection functions for SSD?

[rafisef]

@houseroad I know this has been asked before, but any updates on these models? Any plans for Mask RCNN?

[houseroad]

Yes, we do, still working on it, just low bandwidth.

[brantPTS]

Thanks for asking about this and for the updates. Any new information on status would be great.

[qigtang]

@houseroad any status update on mask-rcnn export?

[houseroad]

I think @linkerzhang is working on it. Any update?

[kubansal]

Hi @linkerzhang did you get a chance to have a look in the mask-rcnn

[imranparuk]

Hi, any update on this?

[daquexian]

@imranparuk Faster R-CNN, Mask R-CNN, yolov3 and SSD has already been introduced in model zoo.

[imranparuk]

@daquexian what about Retinanet?

I am guessing you need the model to be traceable with pytorch. Currently the Retinanet model isn't traceable due to the BoxList (I'm not 100% sure of this, just looking at the issue board for the maskedrcnn benchmark project)? However if its been solved for the Faster-RCNN and Masked-RCNN, surely it shouldn't be an issue for Retinanet?

[SnowRipple]

@daquexian I've been trying to convert RetinaNet to onnx for couple days now but without success. Is RetinaNet officially supported by ONNX now?

[imranparuk]

@SnowRipple Doesn't seem to be the case, from the looks of things the models need to be traceable by torch jit (this doesn't guarantee it can be converted into onnx I don't think).

It has been accomplished by others for other projects as there are onnx models for papers like Faster-RCNN etc. These models make use of similar classes, however there doesn't seem to be any documentation on how to accomplish this.

If you do however find anything, please let me know!

[SnowRipple]

@imranparuk hopefully it is just a matter of time, since most of the Detectron models are already supported, so the RetinaNet should be too.

I agree that a lot of ops that were created for the sake of other detection models can be used in RetiaNet again, however there are still ops that seem to be specific to RetinaNet only.

Personally I find it surprising that such a popular detection model is not supported yet, whereas other (much older with worse performance) models are.

[azuryl]

@houseroad do you have Retinanet onnx