What number and quality of images are needed for labeling and fine-tuning a pretrained model?

Label Studio Support
1

Hello!

I am new to Label Studio and Layout Parser and would greatly appreciate your help with a couple of questions about image annotation for subsequent model fine-tuning (using scripts from this page: GitHub - Layout-Parser/layout-model-training: The scripts for training Detectron2-based Layout Models on popular layout analysis datasets).

  1. How many images would be sufficient to label in Label Studio to fine-tune a pre-trained model (e.g., Faster R-CNN)?
    My dataset contains approximately 15,000 images. ChatGPT suggests labeling 100–200 images initially and 500+ for better performance. In contrast, Copilot recommends labeling 2,000–3,000 images to start.

  2. Do the quality of images and the number/types of labels per image affect the speed of model fine-tuning?
    My images are in PNG format, RGB color space, and have a resolution of 1800x1200.

I look forward to your response!

Best regards,
Roman