Wow, I'm impressed! Such a unique skill! I've never heard anyone do this! Be able to judge the implementation of an entire group of Github repo's grouped by language! Can you do Japanese repo's next?
Its also ridiculous that I am being downvoted.
Your honesty is appreciated. Facts are welcome too.
> I think they mostly use our code.
What's "our code"? The strongly connected set of open source software that happens to be written in English?
I work in machine learning research and keep up with all of the top papers. I have exited on an ml startup that I founded. Very few coming out of china are top papers. Anyone can publish to arxiv. Measuring whether a model is state of the art is incredibly complex, overfitting to specific datasets is rampant.
Its also ridiculous that I am being downvoted.
USA occasionally has a reputation for being insular, but I'm happy to see you bucking the trend and embracing institutes across the continent of America.
I suppose I'm reading too much into why "an American Institute" gets capitalised but "china" or "chinnese" does not?
> overfitting to specific datasets is rampant.
Are your conclusions the result of having overfit to the software/research in your field you've looked at?
I don't know either way, but forgive me if I sense a tiny bit of bias.
4x Zhejiang University, Hangzhou, China
6x Tencent's Youtu Lab, Shenzhen, China
Before that was released, the previous state of the art was: https://arxiv.org/abs/1904.09117
3x The Chinese University of Hong Kong
1x Tencent AI Lab, Shenzhen, China
In Sintel, a benchmark performed by a German university, both of these also rank higher than the highest-ranked research team including a US citizen. It may not be what people want to hear and/or it maybe an issue of cheating, industry espionage, or global politics, but Chinese research teams ARE world-class at computer vision and have been for at least a year now.
Its also ridiculous that I am being downvoted.
But I'll give you that overfitting is an issue, especially on Sintel.
Still, I did work with selflow myself and it is one of the best. The only issue is with thin objects, but skydio (us startup) is having exactly the same problem.
1. No infrastructure related code with a Chinese language .md file on an American open source site that passes the filters that (a) you have seen it, and (b) that is better in your opinion than what "we" (US?) has
to
2. There isn't a universe of software from China we don't see"