briaai

/

RMBG-2.0

like 568

Follow

BRIA AI 1.65k

Hi Peng Zheng,

Thanks for getting in touch! We truly value the foundation you’ve provided, which has played a significant role in our progress.
While we’d love to share the full benchmark dataset, we're unable to do so as it includes proprietary data. However, we're more than happy to provide the dataset distributions.
Please don’t hesitate to let us know if there’s anything else we can assist with.

Best regards,
Or

Thanks for the reply. Have you ever used publicly available dataset for evaluation? For example, DIS-VD in DIS5K and P3M-500-NP in P3M?

While we unfortunately cannot publish the benchmark on which and with which we developed the model, out of respect for copyright, Or has provided the maximum possible information about the benchmark. In the attached Git, [GitHub Link: https://github.com/Efrat-Taig/RMBG-2.0]
there is a script that allows anyone to easily create a benchmark for their specific use case.
Additionally, I’ve prepared another benchmark for comparison. Again, this is not the original benchmark; it is for a specific use case.
I've also created a comparison script between our previous model and the new model:
https://github.com/Efrat-Taig/RMBG-2.0/blob/main/compare_bria_models.py
I haven't yet compared it to BiRefNet. —it would be great if someone from the community could do so.

I know... Thanks for your detailed explanation!

We can discuss also in our Discord community https://discord.gg/Nxe9YW9zHS

Thanks for the details about the test set ! I have a few questions:

1. Is the test set fully generated using the approach from https://github.com/Efrat-Taig/RMBG-2.0 or is it based on real images also?
2. What's the total size of your dataset?
3. Also the discord link seems broken?
   Many thanks :)

The Full Test set is NOT fully generated.
See info about the dataset here: https://blog.bria.ai/brias-new-state-of-the-art-remove-background-2.0-outperforms-the-competition
and here: https://medium.com/@efrat_37973/rmbg-v2-0-model-the-next-generation-in-background-removal-from-images-628360782397
new Discored link :https://discord.gg/Nxe9YW9zHS

Thanks for the quick response and the link to discord !
I didn't find the actual size of the full dataset, could you provide the count?
And also who did the voting to get this graph?

hi @tdurbor the dataset covered few hounders images.
voting was performed manually.

Hi @origubany thanks for the answer, I'm curious about the details around the voting? Who voted? How where the scores calculated?

The voting was based on blind testing conducted by design and art students ,scoring our benchmark results.

how many votes per image were you able to collect?

3 votes per image.

@tdurbor you are more than welcomed to join our discord community channel https://discord.gg/CCNYhjKT to get the latest updates and insights.