Question About Training Duration for a 50K Custom Car Dataset

[imm-nuclier]

Hi @ZhengPeng7 ,

First of all, thank you for your excellent work on BiRefNet.

I have a question regarding the recommended training duration for a custom dataset.

I have a dataset of approximately 50,000 car images for background removal. Although the dataset contains different types of cars (SUVs, sedans, hatchbacks, coupes, sports cars, cars with doors open, cars with antennas, etc.), all images belong to the same general object category (cars).

In your repository, I noticed that you trained a dataset of around 40,000 images for 200+ epochs.

For my dataset, would you recommend training for 200+ epochs as well, or would 50–100 epochs generally be sufficient since the images are relatively similar and belong to the same object category?

Any recommendations or best practices for choosing the appropriate number of epochs would be greatly appreciated.

Thank you!

[ZhengPeng7]

Hi @imm-nuclier,

Sure. Generally, car segmentation is a relatively easier seg task since the shapes of cars are not extremely complex. Besides, your training set is very large.
I would say that 30-50 epochs would likely be sufficient. If your cases are a bit different, you can adapt more.