OpenVino and Re-Identification accuracy

[Duggy]

Hi,

I am using OpenVino Multi Camera, Multi Tracking demo (with re-identification) - (https://docs.openvinotoolkit.org/2020.1/_demos_python_demos_multi_camera_multi_person_tracking_README.html). I am running retail store videos of entrances. The person comes in and gets and ID but I find that when they exit they get a new ID, i.e. the re-identification is hit and miss at best. Is there any way to correct/increase re-identification accuracy? Change Frame rates or re run models or test other models? Something I am missing?

I am utilizing: person-reidentification-retail-0288.xml (tried person-reidentification-retail-0277.xml as well).

Much appreciated!

[Wan_Intel]

Hi Duggy,

Thanks for reaching out to us.

I have tested Multi-Camera Multi-Target Python Demo with OpenVINO 2021.4 LTS using two videos as inputs with the following Intel's Pre-trained model:

Object Detection model: person-detection-retail-0013

Re-Identification model: person-reidentification-retail-0277

When a person entered the detection area, it was assigned to ID1. When the person re-entered the detection area, it was assigned to ID2. However, within a few seconds, it will auto re-assign to ID1.

Also, I have tested Multi-Camera Multi-Person Python Demo with OpenVINO 2020.3 using one webcam as input with the following Intel's Pre-trained model:

Object Detection model: person-detection-retail-0013

Re-Identification model: person-reidentification-retail-0031

The results are the same as using Multi-Camera Multi-Target Python Demo.

I noticed that you are using Multi-Camera Multi-Person Python Demo with OpenVINO 2020.1. Hence, I strongly recommend you use the Multi-Camera Multi-Target Python Demo with OpenVINO 2021.4 LTS which provides bug fixes, longer-term support and maintenance, new capabilities, and performance improvements.

The latest version of the OpenVINO toolkit is available for download on the following page:

https://software.intel.com/content/www/us/en/develop/tools/openvino-toolkit/download.html

Regards,

Wan

[Duggy]

Hi,

Sorry, yes I am using the latest version (just pasted the wrong link). 

Is there any way to "strengthen" re-identification for my use? I am using CCTV cameras do an using from a higher angle, not head on on the same plane as the person, mine tends to be at 45 degrees from the top. Is there any way to "train" the model for my environment but not have to rebuild openVINO?

[Wan_Intel · Accepted solution]

Hi Duggy,

Multi Camera Multi Target demo supports the following object re-identification models for person re-identification:

• person-reidentification-retail-0277
• person-reidentification-retail-0286
• person-reidentification-retail-2087
• person-reidentification-retail-0288

For your information, these person re-identification models use a whole-body image as an input. The pose coverage for these models is standing upright, or parallel to the image plane.

However, I tried running the above models with a camera at 45 degrees from the top using Multi-Camera Multi-Target demo and they are working fine.

On another note, we regret to inform you that these person re-identification models are not available in OpenVINO™ Training Extensions to re-train.

Models that are available to re-train using OpenVINO™ Training Extensions is available at the following page:

https://github.com/openvinotoolkit/training_extensions#models

Furthermore, Deep Object Reid is a library for deep-learning image classification and object re-identification, written in PyTorch. It is a part of OpenVINO™ Training Extensions.

For more information on Deep Object Reid, you may open the ticket on the following page:

https://github.com/KaiyangZhou/deep-person-reid/issues

Regards,

Wan

[Wan_Intel]

Hi Duggy,

This thread will no longer be monitored since this issue has been resolved. 

If you need any additional information from Intel, please submit a new question.

Regards,

Wan