top-cvpr-2023-papers: a curated list of CVPR 2023 papers with code

What top-cvpr-2023-papers actually solves

CVPR is large enough that the accepted-papers list is not a reading list. The README states that 9,155 papers were submitted to CVPR 2023 and 2,359 were accepted. Nobody triages that by scrolling. This repository compresses the conference into a table of papers the maintainer considers worth attention, and it does the one thing a plain accepted-papers page does not: it pairs each entry with a code link where one exists, so you can move from a title to a repository in a single click.

The audience is narrow and specific. It is for a computer-vision engineer or graduate student who wants to know which 2023 papers have runnable code and fall into areas like segmentation, 3D reconstruction, NeRF, generative image models, or benchmarking. It is not for someone who needs the full proceedings, a BibTeX export, or per-paper citation counts. The README is explicit about that boundary: if the paper you want is not on the short list, it directs you to the full list of accepted papers on the CVPR site.

The table, the topic column, and the generated-file warning

The mechanism is deliberately boring, which is the right call for a list that has to stay readable in a diff. The README contains one Markdown table with three columns: topic, title, and a repository/paper cell holding a GitHub badge link and an arXiv badge link. Topics in the visible rows include Segmentation, NeRF, 3D, Generative AI, Transfer Learning, and Benchmarking. Some rows carry compound topics such as "Segmentation and Generative AI," which is a useful signal that the topic column is a human label rather than a taxonomy with controlled vocabulary.

The important structural detail sits above the table. HTML comments mark the region as AUTOGENERATED_COURSES_TABLE and warn: "DO NOT EDIT THIS TABLE MANUALLY. IT IS AUTOMATICALLY GENERATED." The warning then points to CONTRIBUTING.md for the correct way to make changes. That single comment tells you most of what you need to know about the project's architecture. The table is an output. Whatever you see in the README was produced by something else, and the repository layout confirms it: alongside README.md, LICENSE, and .gitignore, there is an automation/ directory. The natural reading is that scripts in automation/ regenerate the table, and that a contribution is a change to the source data or the generator, not a hand-edit of the Markdown. The README does not spell out the generator's input format, so the automation/ directory is where you would look to confirm it.

One consequence is worth stating plainly. Because the table is generated, the ordering and the inclusion decisions are only as current as the last run. The last push to the repository was on 2026-04-15, so the README as it stands reflects whatever the generator produced at that point.

Reading the table without installing anything

There is no package to install and no CLI to run. The README gives no install steps because there is nothing to install: the deliverable is README.md itself, and the README does not document any download command, environment variable, or configuration key. The only practical step is to open the file on GitHub or clone the repository, which is the repository URL shown in the README and on the project page.

Once you have the file, the table is plain Markdown, so any text search over it works. The topic column is the fastest way to narrow the list to one area, and the third column carries the two links you actually need: a GitHub badge for the code and an arXiv badge for the paper.

What you should expect to see in a row is a topic label, a title, and a cell containing both badges. What you should not expect is a PDF, an abstract, a citation key, or a per-paper summary. The README does not host paper content, and it does not provide a download of the papers themselves.

If you want to propose an addition, the README's own instruction is to read CONTRIBUTING.md rather than editing the table. The README does not document what that file contains, so treat it as the entry point for the contribution process rather than a known procedure.

Where the curation breaks down

The biggest gap is selection. The README explains why the list exists and how large the conference was, but it does not state the criteria for inclusion, who reviews a proposed entry, or how many papers are on the list. A curated list without published criteria is a list you have to trust. That is a real limitation for anyone using it to decide where to spend a week of reading time, because you cannot tell whether an omission means the paper was judged weak or simply was not seen.

There is also no freshness guarantee on the links. The table points to third-party arXiv entries and third-party GitHub repositories. Those repositories are maintained by the paper authors, and nothing in this project tracks whether they still build, whether dependencies have rotted, or whether a repository has been archived. A code badge in this table means a repository existed when the entry was added, not that the code runs today.

Finally, the list is scoped to 2023 by design. The README links to sibling repositories for 2024, 2025, and 2026, which is a clean way to handle the year boundary, but it also means this repository will not absorb later work. If your interest is current, this is the wrong entry point. And if you need the complete record, this is the wrong tool entirely: the README itself sends you to the official accepted-papers list for anything outside the short list.

How it compares with the official accepted-papers list and with Papers with Code

The official CVPR accepted-papers page is the authoritative source and the one this README defers to. Its approach is exhaustive and unranked: it lists accepted papers without editorial judgement and, in the form the README links to, without code links. That is exactly what you want for a literature review or a completeness check, and exactly what makes it unusable for triage. The difference here is not data quality, it is the presence of a human filter and a code column.

Papers with Code takes a different route to the same problem. It indexes papers at scale and attaches code, datasets, and benchmark leaderboards, which means coverage is broad and machine-driven rather than curated. The trade-off is the inverse of this repository's: you get far more papers and far less signal about which ones a practitioner thought were worth reading. A curated short list with code links and a broad index with leaderboards solve different tasks, and neither replaces the other.

There is a third comparison worth making, and it is the honest one. For many questions, a plain web search or an arXiv listing filtered by conference will get you to the same papers. What this repository adds over that is the pairing of title, topic label, and code link in one static file that you can clone and search offline.

Licence, maintenance, and what upgrading costs you

The repository is licensed CC0-1.0, which places it in the public domain to the extent the licensor can do so. In practice that means you can copy the table, rehost it, or fold it into internal documentation without an attribution obligation. Two caveats follow from the structure rather than the licence text. The licence covers this repository's own content, not the papers or the linked code repositories, which carry their own terms. And a CC0 dedication on a curated list does not grant any rights in the abstracts or figures it might point to; the table links out rather than reproducing paper content, which keeps that boundary clean. This is a description of what the licence file states, not legal advice.

The last push to the repository was on 2026-04-15, so the README has not changed in roughly five months. That is consistent with what the project is: a 2023 conference list does not need frequent edits once the conference is over, and the sibling repositories for later years absorb new material. The upgrade cost is therefore close to zero for a consumer. You clone it, you read it, and if it is updated later you pull. The cost sits on the contribution side instead: because the table is generated, a change means understanding the generator in automation/ and the process in CONTRIBUTING.md, which is more work than editing a Markdown row. That is the price of keeping the table consistent, and it is a reasonable trade for a file this size.

Editorial conclusion

Use it if you want a short, code-linked shortlist of CVPR 2023 work and you accept that the selection criteria are not documented. Do not use it as a citation source, as a complete conference index, or as a way to reproduce results: the README points to the official accepted-papers list for completeness and says nothing about how entries are chosen. Before relying on it, open the automation/ directory and CONTRIBUTING.md on the master branch to see how the table is generated and how a change gets merged.

Frequently asked questions

Official sources

1. Issues
2. License: CC0-1.0
3. README
4. SkalskiP/top-cvpr-2023-papers on GitHub