Updating pre-trained YOLO v4 weights with new images for custom data-set

[v-nayjack]

@AlexeyAB @WongKinYiu

I have a pre-trained weights file for a custom object detection. The number of images available for this custom data-set is very small, however I add more and more images as and when they become available to me.

The objects I am trying to detect are very small compared to the input image, so I decided to calculate the anchors based on the custom data-set I created and use it for training the YOLO v4 model. After the initial training, I want to understand the following things:

1. Can I just add new images to my data-set and use the xx_final.weights to retrain/update the YOLO v4 model? How many batches do I need to retrain the model? e.g. For 50 classes, my initial training ran for 100k batches, after adding the new images to the data-set, do I need to retrain the model for 100k batches again? Also, I am including old & new images every time I add new images to my custom data-set and retrain the model.
2. Should I use same anchors as initial training for adding new images to the training?
3. Is it okay to change the anchor values? e.g. I'll add the new images to the same data-set and recalculate the anchors and update the .cfg file with these new anchor values (In this case I am keeping the network size same).
4. How would changing the anchor values affect the object detection for the "updated" training?
5. E.g., If I calculated the anchors for 416x416 and trained the model, Can I change the network size to 608x608, recalculate anchors (since I'll be adding more images to my custom data-set) and "retrain/update" the training using the xx_final.weights file?

Any help is greatly appreciated!

Thank you in advance,

-VJ
#7869