MV Sense custom CV models

SOLVED
PeterAC
Conversationalist

MV Sense custom CV models

Hi all,

 

We are analyzing the possibilities of object detection with custom computer vision models in our Meraki MV12WE, and we are finding some problems. We have uploaded two custom models, overcoming the related restrictions of size, model format and number of outputs, but We don't get any detections when running them. One of them is the ssd-mobilenet-v1.tflite from the official meraki website, another is an efficientdet-lite0 model trained according to Google's indications for tensorflow lite models. As I said, in both cases boxes are not drawn when displaying objects in stored recordings, We don't know if these results are only exported by MQTT, or that We are missing steps to take into account to test this functionality (we assume that the mobilenet model provided by Meraki does not have output order problems).

 

We appreciate any contribution in this regard, thanks and best regards.

1 ACCEPTED SOLUTION
PhilipDAth
Kind of a big deal
Kind of a big deal

I don't know the answer.

 

My vague recollection is this was only for use by MQTT.  Search for MQTT in this document (three occurrences):

https://documentation.meraki.com/MV/Video_Analytics/MV_Sense_Custom_Computer_Vision 

View solution in original post

5 REPLIES 5
PhilipDAth
Kind of a big deal
Kind of a big deal

I don't know the answer.

 

My vague recollection is this was only for use by MQTT.  Search for MQTT in this document (three occurrences):

https://documentation.meraki.com/MV/Video_Analytics/MV_Sense_Custom_Computer_Vision 

Thanks for your answer, We will try that functionality.

KrisPL
Conversationalist

Hey! I had same issue, finally I've trained mobile ssd_mobilenet_v2 with tf 1.15 version, and it worked unfortunately it was not possible to train efficient det, but it may be if you will change hyper parameters in terms of the input and output from the network. The TF version needs match to the version supported by the camera gen, you need to check it first. Good luck!

PeterAC
Conversationalist

Thanks KrisPL. I will try with mobilenet v2, I've trained some efficentdet models in Google Colab based on this tutorial. Standard EfficentDet models are compatible with the camera and return detections by MQTT but very inaccurate. Trained ones don't return any detection for the moment. I need to harvest new data.

KrisPL
Conversationalist

Hi! I did the same in the first run, this tutorial is using TF 2.8 which is not supported by the device, the highest supported version of the TFlite is 2.3 on gen2 cameras and 2.5 on gen3 (it didn't worked for me), and standard EFDet outputs are different than Meraki expected. Check this docs: model training: https://developer.cisco.com/codeexchange/github/repo/Francisco-1088/merakiCustomCvDemo , model conversion https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_on_mobile_t... , about Outputs: https://www.tensorflow.org/lite/examples/object_detection/overview.

 

Let me know if you succeed!

 

br, Kris

Get notified when there are additional replies to this discussion.
Welcome to the Meraki Community!
To start contributing, simply sign in with your Cisco account. If you don't yet have a Cisco account, you can sign up.