How to Update an Object Detection Network Created with Mask R-CNN and Detectron2

[wodud4143]

Hello, I am a student studying Vision AI. Currently, I have created a network to detect food items on a tray using Mask R-CNN and Detectron2. However, the problem is that it cannot detect food items that it has not been trained on. I have managed to crop the undetected food items from the images, but I want to retrain the model with these undetected images to update it. The existing model was created by labeling images with Labelme, converting them to the COCO dataset format, and using the model zoo. My planned method for retraining (i.e., updating) the model is as follows:

1. Convert the undetected food item images into Labelme JSON format.
2. Convert them to the COCO dataset format.
3. Merge the annotations created in step 2 with the existing annotations.
4. Add the new images (undetected food item images) to the existing image folder.
5. Retrain using the combined annotations and updated image folder (model update).

Is this theoretically possible? Additionally, I want to retrain the model so that it learns less about the existing classes and more about the new classes. If there is a better way to update the model, please let me know.

I want the model to learn from the undetected food item images so that it continues to detect the previously detected food items correctly while also detecting the newly learned food items accurately.