Yolo models don't use any historical data for inference. In fact, YOLO is an acronym for You Only Look Once. Generally, I find that adding some hysteresis filtering to the bounding boxes in postprocesing gives you the smoothness you need. You can see what we do in our demos here. If you run your model through rpicam-apps, you can enable this filtering in your json config file like we do here.
Other things to try out would be to use Yolov8 to see if that give you better results, or perhaps try adding augmentation to your dataset to improve performance when objects are bigger/smaller or rotated.
Other things to try out would be to use Yolov8 to see if that give you better results, or perhaps try adding augmentation to your dataset to improve performance when objects are bigger/smaller or rotated.
Statistics: Posted by naushir — Thu Oct 30, 2025 8:55 am