Show HN: CVPR 2020 unofficial statistics, with better search functionality | Hacker News Reader

Show HN: CVPR 2020 unofficial statistics, with better search functionality

cvpr20.janruettinger.com

40 pointsck_one13 comments

CVPR 2020 statistics (unofficial) + better search functionality

tldr; I have created a dataset about CVPR 2020 papers consisting of the title, author(s), affiliated institution(s) and the abstract of each paper and put it behind Elastic Search to make it more accessible. Happy searching!

Initially, I wanted to find out which research institution is in involved in what papers. To my surprise this information wasn't readily available. I had to go through each of the 1500 papers to extract the information. I used a script to get the title, author(s) and the abstract of each paper and worked with a freelancer (100$/~30h) to get the institution of every author. Then I used local sensitive hashing to clean institution names and put the whole dataset behind Elastic Search. A simple idea turned out to be a good learning project since it was my first time working with a freelancer and also my first time using Elastic Search.

Quick summary about CVPR 2020 statistics: Google is still the number one in terms of number of publications. China is gaining momentum quicker than I expected.

The blank page without context gives no clues as to what words might make interesting searches. Maybe put a list or word cloud of key words off to the side to suggest ideas?

Good point! Here is a list of keywords I have used: ["Oxford", "Technical University of Munich", "Self Supervised", "3D Reconstruction"]. Will add a list of keywords after dinner on the site.

Wow, I'd noticed the increase in Chinese sounding author names during the past few years but had no idea China was such a major player.

Glad to see TUM being represented in this with a couple of papers as well...

TUM == Technical University of Munich. TUM has a strong Computer Vision department with Niessner, Leal-Taixé and Cremers.

Great project! Obviously a stretch, but would be curious to see a breakdown by workshop, general conference, orals, etc. The umber of publications is not the same as their quality.

Also, do you disambiguate corporate affiliations in academia? E.g. many papers will be +1's for a university and a corporate research group (Stanford + Google)

That would be interesting indeed. The way I do it right now is that every institution listed on a paper gets +1 submission for that paper. That's not always fair since sometimes corporate labs only provide resources and no other contributions. But I couldn't come up with a good method to resolve this issue. I think the numbers are a good approximation of reality.

Thanks for the great work!

Btw If I may ask, when you compiled the Institutions, did you apply some disambiguation? For example consider that UCLA === University of California, Los Angeles

Great question! A combination of techniques were involved in cleaning up institution names. First, I deleted all department names so that only the university/lab name was left. Second, I used local sensitive hashing (wiki: https://en.wikipedia.org/wiki/Locality-sensitive_hashing) to find similar strings. LSH maps similar strings to the same bucket. This resolved some ambiguity like: UC Berkeley/University of California Berkeley etc. But there was also a good chunk of manual clean up necessary at the end.

This is the traffic I got following the HN post:

- total number of search requests: 3500

- ~20 search requests/min (at the moment)

I never had that much traffic with a private project. Thanks for your interest!

Great initiative! How long did it take you to parse through all the papers etc and build this tool?

The coding part of the project (extracting information, setup elastic search, build website) took me around two days. But some information had to be extracted manually. I hired a freelancer on upwork for that who spent ~30hs on it. It would be cool if conferences would collect data directly with the submission more carefully so that we can analyse more easily which country/lab is getting stronger and why.

Reply on news.ycombinator.com