I’m currently working with the hailo_apps_infra repository for object detection and tracking using GStreamer.
In my pipeline, I’m using the TRACKER_PIPELINE to integrate hailotracker, set to track all object classes by using class_id = -1. During inference, the tracker assigns IDs correctly, but I’ve noticed that the FPS drops significantly within a few seconds when tracking across all classes. Interestingly, this performance drop doesn’t occur when restricting tracking to a single class. Do you have any insight into why this might be happening?
Also, in the PyTorch implementation of YOLOv8, the BoT-SORT tracker is used and performs consistently with minimal ID switching. Is it possible to integrate BoT-SORT (or a similar external tracker) into the Hailo GStreamer environment, or is the pipeline limited to hailotracker?
Thanks in advance for your support!