top-cvpr-2024-papers is a generated shortlist with a do-not-edit block and uneven links

The list block is generated, and it says so twice

The papers sit between two markers in the README, one of which is a visible comment aimed at contributors:

<!--- AUTOGENERATED_PAPERS_LIST -->
<!---
   WARNING: DO NOT EDIT THIS LIST MANUALLY. IT IS AUTOMATICALLY GENERATED.
   HEAD OVER TO https://github.com/SkalskiP/top-cvpr-2024-papers/blob/master/CONTRIBUTING.md FOR MORE DETAILS ON HOW TO MAKE CHANGES PROPERLY.
-->

So the useful path for adding or correcting a paper is not a pull request that edits the README, and the warning redirects to CONTRIBUTING.md on the master branch instead. The repository root is small enough to see the shape of the thing: .gitignore, CONTRIBUTING.md, LICENSE, README.md and an automation directory. That directory is the machinery behind the block, and it is also the only reason the repository's primary language is Python, since there is nothing here for an end user to install or run. Every entry in the list was placed by that process rather than typed in.

The short list is a fraction of 2,719 accepted papers

The scale is stated up front. In 2024 alone, 11,532 papers were submitted and 2,719 were accepted, and this repository is the author's answer to the question of how anyone is supposed to find the good ones inside that number. The framing is explicit that it is a short list, and it names the escape hatch for everything else: the conference's own accepted papers page. Read that way, the repository is triage rather than coverage. What it adds over the official page is not the titles but the packaging, because each entry carries a poster image, an arXiv link, a code link, a demo, a video or colab entry, an author list, a topic label and a session slot, gathered into one place per paper. The topic labels used across these entries include 3d from multi-view and sensors, deep learning architectures and techniques, document analysis and understanding, and efficient and scalable vision.

Every entry links a different subset of paper, code, video, demo

The link payloads are not uniform, and that is the main thing to know before you start clicking. SpatialTracker, marked as highlighted, carries a poster, a paper and code. ViewDiff carries a poster, a paper, code and a video. OmniGlue carries a paper, code and a demo but no poster image. DocRes has a paper, code and a demo, also without a poster. The clearest data defect among these entries sits in Florence-2, which is highlighted and otherwise well supplied with a poster, a paper, a video, a demo and a colab entry: the colab link and the video link point at the same youtu.be address, listed twice under two different labels. So a link count is not a quality signal here, and the same URL under two labels is proof that the fields are filled in per entry rather than derived from one record.

Poster images ride a timestamped CVPR media URL

When an entry has a poster, the image comes from the conference's own media path, in a shape like PosterPDFs/CVPR%202024/31668.png followed by a query string of the form t= followed by a long decimal number. Those numbers are Unix timestamps and they differ from poster to poster, so 31668 and 31616 and 30529 and 30022 arrive with different t values that reflect when the file was fetched rather than anything about the paper. The entries without a poster anchor show that the field is optional in practice, which is worth noting before assuming every paper here has a poster page to look at. What stays constant across the sample is the ID in the path, which is the paper's poster number inside the conference's media store.

A flame mark picks out the highlighted entries

Highlighted papers are marked rather than separated, using a flame glyph placed in front of the title in the entry's bold heading. Among the entries on this page, the marked titles are SpatialTracker: Tracking Any 2D Pixels in 3D Space, Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks, and EfficientSAM: Leveraged Masked Image Pretraining for Efficient Segment Anything. Everything around the mark is ordinary entry markup, so a reader scanning for it sees three signals at once: poster, arXiv, and the title, followed by authors and links. Worth noting where those marks sit, because two of the three land on papers with a Hugging Face demo and the third one has a video as well, while the unmarked entries in the same topics often have only a paper link.

Demos point at third-party spaces, except the one the curator hosts

Demo links go to Hugging Face Spaces, and the visible sample shows four different hosting arrangements. OmniGlue's demo sits under a qubvel-hf account, DocRes's demo sits under the same qubvel-hf account, and Florence-2's demo sits under gokaygokay. EfficientSAM's demo is the exception: it lives under a SkalskiP space, the same name as the person maintaining these curated lists. So most demos are third-party deployments assembled by whoever shipped the model, while at least one is a deployment the collection's author put up. That distinction is invisible from the entry itself, since the link label is the same word in all four cases, and a Space is a hosted application that can be asleep, rate limited or removed after a conference year.

Session lines carry poster numbers and time windows from the event

Each entry ends with a session line, and that is the part of the record that dates fastest. The visible slots look like Fri 21 Jun 1:30 p.m. to 3 p.m. EDT with poster number 84, Wed 19 Jun 8 p.m. to 9:30 p.m. EDT with number 20, another Wed 19 Jun evening slot at number 32, a Wed 19 Jun slot at 102, a Thu 20 Jun slot at 101 and another at 144. Numbers are not sequential and are not grouped by topic, so they are the conference's own poster identifiers. What the repository's last push date, 2026-04-15, means is that these rows describe a schedule that has already run. There are no GitHub releases either, so no tagged version pins the state of the list at the moment the conference happened.

Editorial conclusion

Use this repository to triage a conference you did not attend, not as a bibliography. It is a set of pointers into poster pages, arXiv, code repositories and Hugging Face demos, grouped by topic, and the session times it records belong to a conference that has already happened. Before you rely on an entry, open the linked repository rather than trusting the presence of a link, since code and demo availability differs paper by paper, and check CONTRIBUTING.md if you want an entry changed, because the list itself is generated.

Frequently asked questions

Official sources

1. Issues
2. License: CC0-1.0
3. README
4. SkalskiP/top-cvpr-2024-papers on GitHub