openvinotoolkit
Multi Camera Multi Tracking improving Reid accuracy
Hi,
Is there a way to improve the Reid accuracy? When we deploy to different sites we never have the same angles, the same lighting conditions (even between cameras). Is there any tool/solution to improve the model reidentification accuracy or the person identification accuracy per site or per camera? We get a lot of chairs being identified as people, reidentification drops and reconnects (sometimes) and between cameras (angles, lighting conditions etc) reid isn't fantastic (cannot track a person between cameras reliably).
Wondering if there was a solution or tool that we can run per site, per camera in order to improve reid and person identification accuracy.
Much appreciated.
@duggydoo the demo use person-detection-retail-0013 model for object detection. The model was trained on certain dataset and expect objects in specific view angle, lighting conditions and so on. To fine tune model for your specific case usually training is required on your specific data. There is OpenVINO Training Extension, which support fine-tuning for some of Intel models.
Thanks, but I take it no way to improve reidentification model for specific settings? Only person detection?
@duggydoo I see no reidentification model support on OpenVINO Training Extensions repo, you may ask about plans for reid support directly on their github repository.
cc @snosov1
