Source of Retinaface model and accuracy metrics?

Hi Team,

Thank you for releasing this great project. I've a got a few questions about the RetinaFace face detector and would really appreciate your response.

There are two RetinaFace papers.

[1](https://arxiv.org/pdf/1905.00641.pdf). RetinaFace: Single-stage Dense Face Localisation in the Wild
[2](https://openaccess.thecvf.com/content_CVPR_2020/papers/Deng_RetinaFace_Single-Shot_Multi-Level_Face_Localisation_in_the_Wild_CVPR_2020_paper.pdf). RetinaFace: Single-shot Multi-level Face Localisation in the Wild

- Both of these papers have different accuracy on WiderFace listed
- Retinaface's README shows a different accuracy of [`WiderFace validation mAP: Easy 96.5, Medium 95.6, Hard 90.4.`](https://github.com/deepinsight/insightface/tree/afda338f887bb766ae0b0ee0deeebaddeb1d31dd/detection/retinaface#:~:text=WiderFace%20validation%20mAP%3A%20Easy%2096.5%2C%20Medium%2095.6%2C%20Hard%2090.4.)
- SCRFD's README shows a different accuracy of [`RetinaFace (CVPR20) ResNet50 94.92 91.90 64.17 29.50 37.59 21.7`](https://github.com/deepinsight/insightface/tree/afda338f887bb766ae0b0ee0deeebaddeb1d31dd/detection/scrfd#:~:text=RetinaFace%20(CVPR20),21.7)

Four questions here:
1. Why is accuracy different for `RetinaFace-ResNet50` in [RetinaFace's README](https://github.com/deepinsight/insightface/tree/afda338f887bb766ae0b0ee0deeebaddeb1d31dd/detection/retinaface#:~:text=WiderFace%20validation%20mAP%3A%20Easy%2096.5%2C%20Medium%2095.6%2C%20Hard%2090.4.) and [SCRFD's README](https://github.com/deepinsight/insightface/tree/afda338f887bb766ae0b0ee0deeebaddeb1d31dd/detection/scrfd#:~:text=RetinaFace%20(CVPR20),21.7)?
2. From where is the Retinaface's README accuracy of [`WiderFace validation mAP: Easy 96.5, Medium 95.6, Hard 90.4.`](https://github.com/deepinsight/insightface/tree/afda338f887bb766ae0b0ee0deeebaddeb1d31dd/detection/retinaface#:~:text=WiderFace%20validation%20mAP%3A%20Easy%2096.5%2C%20Medium%2095.6%2C%20Hard%2090.4.) sourced?
3. [RetinaFace-R50 pretrained mode](https://github.com/deepinsight/insightface/tree/afda338f887bb766ae0b0ee0deeebaddeb1d31dd/detection/retinaface#:~:text=Pretrained%20Model%3A%20RetinaFace%2DR50%20(baidu%20cloud%20or%20googledrive) ) is from which paper?
4. Is there any chance that improved models will be released?

Could you please clarify these things?

Thanks again for your response and contributing to the open source community.