Source Model for person-detection-retail-0013

[Sandhu__Ranjit]

Hi There,

I am looking for source mode "caffe" as mentioned in deployment_tools/intel_models/person-detection-retail-0013/description/person-detection-retail-0013.html. 

I have downloaded models using Model_downloader but it gives me  Retail\object_detection\pedestrian\rmnet_ssd\0013\dldt\person-detection-retail-0013-fp16.xml and  person-detection-retail-0013-fp16\bin files and not .cafee model file.

I want to use this model with model optimizer to customize its behaviour.

Thanks in advance.

~Ranjit

[Sandhu__Ranjit]

Hi there ,

it looks like  the pre trained model used int for "person-detection-retail-0013" is  "rmnet_lrelu_pd_ssd.caffemodel" . So is it possible to get the model ?

Any help will be much appreciated

~Ranjit

[Sandhu__Ranjit]

Hi Intel,

Please get me some direction on my question?

Regards,

Ranjit

[Dmitry_K_Intel3]

Check https://github.com/opencv/open_model_zoo/tree/2018/intel_models/person-detection-retail-0013. There is a Caffe's prototxt.

> I want to use this model with model optimizer to customize its behaviour.

Please clarify what indeed should be changed. Maybe it's possible without Model Optimizer.

[Sandhu__Ranjit]

Hi Dmitry,

Thanks for answering my query. 

I want detection process to be really fast which will further  make tracking fast. For this to happen, I think if model be restricted to the  detection of persons having fixed size( probably bigger size) then all the persons which far away from the pedestrian crossing will not be detected as those appear to be of smaller in size as compare to actual people crossing the pedestrian.

So this way we have fewer detections/tracking which will ultimately improve its performance.

If you look at the current behaviour it do recognize the persons which are not really pedestrians and are far away from the pedestrain crossing but still it consider as pedestrians.

Appreciate any help and suggestions to make it fast.

Regards,

Ranjit

[Gdeep]

hello sandhu saab,

Tusi pedestrian accuracy probability flag '-t' di value increase ker k deko, it will work fine   

[Sandhu__Ranjit]

Thanks Gagan!

I tried but no luck.

~ranjit

[Gdeep]

When try with your own data-set to train your custom model

[Sandhu__Ranjit]

Sorry, could not get you. 

This is what I did as per your suggestion , I ran sample example of Pedestrian_tracker_demo example by setting -t to 100 as command line parameter and did not affect the outcome.

~ranjit

[Gdeep]

first, you just check what the parameters we need to pass to run pedestrian_tracker_demo, I thing -t flag is not there, and the value of probability check flag should be float, not an integer, -t pass 0.90 

[Sandhu__Ranjit]

Yep your are right 't' s not an option for this model but I tried as you suggested :). Surprisingly running a sample with wrong optional argument does not indicate any error about wrong argument.

Any other way to achieve the same thing ?

~ranjit 

[Sandhu__Ranjit]

Nale veere apni bhasha vich nahi likh reha kyu mera hor colleagues vi mainu follow kar rahe ne is problem layee ise forum te  . Hope u understand :)

[Gdeep]

see these are pre-trained samples so you can use these samples, and if you want more speed of inference then you can go with tiny yolo, 

and you from Punjab?

[Sandhu__Ranjit]

Ok will spend some more time to your suggestions and will let you the results.

Yes, I am from Punjab but now settled in Australia.

~ranjit

[Gdeep]

For the better understanding of your problem may I know what actually you want to do? 

[Sandhu__Ranjit]

As mentioned earlier

"

I want detection process to be really fast which will further  make tracking fast. For this to happen, I think if model be restricted to the  detection of persons having fixed size( probably bigger size) then all the persons which far away from the pedestrian crossing will not be detected as those appear to be of smaller in size as compare to actual people crossing the pedestrian.

So this way we have fewer detections/tracking which will ultimately improve its performance.

If you look at the current behaviour it do recognize the persons which are not really pedestrians and are far away from the pedestrain crossing but still it consider as pedestrians."

So i want to use this model t0 count number of  people passed from one gate and keep tracking till they  go out of sight .

Ranjit

[Dmitry_K_Intel3]

If you look at the current behaviour it do recognize the persons which are not really pedestrians and are far away from the pedestrain crossing but still it consider as pedestrians.

 @Sandhu, Ranjit, Please provide some examples with images. And it's still a bit unclear for me, do you want to improve an accuracy of detections or efficiency of model (reduce inference time)?

[Mitra__Debasish]

@Dmitry: I would like to access the cafee model file too so that I can benchmark and compare performance vs speed improvement for regular model vs optimized IR model.