openvinotoolkit / open_model_zoo

Pre-trained Deep Learning models and demos (high quality and extremely fast)
https://docs.openvino.ai/latest/model_zoo.html
Apache License 2.0
4.1k stars 1.38k forks source link

face re-identification limits #3166

Closed yahanda closed 2 years ago

yahanda commented 2 years ago

Hi there!

I have a quick question about maximum number of person that can be registered in face gallery for face-reidentification-retail-0095. Are there any limits? Any known usage restrictions?

vladimir-dudnik commented 2 years ago

Looking at face-reidentification-retail-0095 model description, we can see that it takes image with one face (128x128) at a time and produce vector of 256 floating point values, representing descriptor of this face. This descriptor then may be compared by cosine distance against such vectors, computed on other images, potentially representing other faces (or different views of the same face) to determine if they are similar enough. So, it is not a model but rather application responsibility to store as much descriptors as needed. The limiting factors, application developers need to take into account, are performance of comparison all descriptors. For offline application, obviously there are almost no limits (probably except amount of memory available at system). For real-time applications. limiting factor is time required for comparison all descriptors. But again, implementation of cosine distance and comparison of descriptors by cosine distance done at application level, it is not a part of model inference, not a part of OpenVINO runtime.

yahanda commented 2 years ago

Understood. Thank you for the clarification.