Discrepancy in Annotation Decimal Precision & Partial Dataset Download in Roboflow

Community Help
1

Hi everyone,

I’m currently using Roboflow to generate YOLO format annotations, and I have a few questions.

Firstly, I’ve noticed that the decimal precision of bounding box coordinates in Roboflow is significantly higher than in other platforms like CVAT. For instance, Roboflow often outputs coordinates with up to 17 decimal places, while CVAT typically uses around 6 or 7.

Here are my questions.

  1. What is the rationale behind Roboflow’s higher decimal precision?
    Does it lead to significantly better model performance, especially for smaller objects or high-resolution images?

  2. Is it okay to combine datasets with different levels of precision (e.g., Roboflow vs. other platforms) for training? Would this cause any issues in the training process or model performance?

  3. If combining datasets with different precision levels isn’t advisable, is it a good practice to reduce the precision of Roboflow annotations (e.g., truncating to 6 or 7 decimal places) to match other datasets? Would this affect the model’s performance?

Additionally, I’m wondering if there’s a way to partially download and add files to an existing dataset in Roboflow, rather than re-downloading the entire dataset. This would be really helpful when I only need to update or add a small number of images and their annotations.

Any insights or suggestions would be greatly appreciated!

Thanks in advance for your help!

2

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.