Is There Edgetpu Tflite Model Based On Open Images for Object Detection?

[jsl303]

Is there any object detection model for coral based on open image dataset >= v4?
I tried to convert the models below with no luck.I wonder if anyone already has.

• faster_rcnn_inception_resnet_v2_atrous_oidv4
• ssd_mobilenetv2_oidv4
• ssd_resnet_101_fpn_oidv4

https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md

It would be nice to be able to detect more than 90 classes from coco.

[Namburger]

@jsl303 we only had canned object detection models with the coco dataset, although we do have tutorials for transfer learning with your own dataset.
For the models that failed to converts, have you got a chance to check if it meets the requirements?

[jsl303]

I looked at the requirements, but unfortunately it's beyond my head. If the 3 mentioned models from Tensorflow model zoo meet the requirements, it would be awesome if someone with expertise could convert and include them in this repo. Of course, transfer learning is certainly a nice option, but I think it would be amazing for newbies like myself to be able to run object detection over 600 categories with Coral out of the box.

[jsl303]

It looks like PINTO0309/PINTO_model_zoo was able to convert and quantize to full int8 and compile for Coral Edge TPU.

PINTO0309/PINTO_model_zoo#13 (comment)