Intel® Distribution of OpenVINO™ Toolkit
Community assistance about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all aspects of computer vision-related on Intel® platforms.
6501 Discussions

Improve reid on multi camera multi tracking demo per camera per environment

Duggy
New Contributor I
2,052 Views

Hi,

 

Is it possible to improve the model of Reid/person identification per environment, per camera? When we deploy in a given environment or on a given camera, the conditions are not always ideal which impacts Reid accuracy (be it angle, lighting, view etc). Is there a way to train or improve the accuracy of Reid per camera, per environment? 

 

Thanks

0 Kudos
5 Replies
Peh_Intel
Moderator
2,022 Views

Hi Duggy,


Referring back to the previous discussion, changing the time_window to 1 can increase the accuracy and reducing track_clear_thresh to 300 or less to prevent the accumulation of tracks.


If you wish to train your own model, you may refer to Deep Object Reid, which is a library for deep-learning image classification and object re-identification, written in PyTorch.

 

 

Regards,

Peh


0 Kudos
Duggy
New Contributor I
2,018 Views
0 Kudos
Duggy
New Contributor I
2,012 Views

Hi,

 

This solution would be to replace the reid model, not necessarily to fix the tracker being overwhelmed issue? In order to correct that we need to rewrite that section and somehow pull it out and do it in parallel as to not effect system performance. Is this correct? 

0 Kudos
Peh_Intel
Moderator
1,993 Views

Hi Duggy,


The existing Multi Camera Multi Target Python Demo supports Intel Pre-trained reidentification models: person-reidentification-retail-0277, person-reidentification-retail-0286, person-reidentification-retail-0287, person-reidentification-retail-0288. Unfortunately, in your previous post, you claimed that you obtained unsatisfied performance when running the demo with the Intel Pre-trained model in your scenario.


Hence, train your own model based on your scenario to replace the Intel Pre-trained model. Theoretically, it should be better. However, there are also many further challenges such as huge number of own datasets, consumed time in training model, compatibility of own model with the existing demo, and even the performance also.

 

 

Regards,

Peh


0 Kudos
Peh_Intel
Moderator
1,972 Views

Hi Duggy,


This thread will no longer be monitored since we have provided suggestions. If you need any additional information from Intel, please submit a new question. 



Regards,

Peh


0 Kudos
Reply