I would like to have a capability of my object detection model to improve over time as the available dataset grows over time, but avoid re-training a model from scratch every time I want to update the model.

To be concrete, I am using a YOLOv4 based on Darknet to train a custom set of weights given a custom dataset. Right now if my dataset grows 10%, the only practical way of incorporating "new knowledge" into the model is to re-train it from scratch on a whole dataset. Is there a technique, perhaps outside of YOLO in alternative object detection systems, where "additional" training is possible, perhaps with a reduced learning rate for new data?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.