Cvpr Related Work

CVPR Related Work

At CVPR the related-work problem is not scarcity but velocity: the field's literature updates weekly on arXiv, the venue itself accepted 4,090 papers in the 2026 edition alone, and reviewers are drawn from a 25,000-person pool that includes the authors of whatever you forgot to cite. Position precisely or be positioned by Reviewer 2.

Five shelves to cover

The nearest-neighbor shelf deserves half the section's words. For each: one sentence of credit, one sentence of mechanism difference, one sentence of consequence ("because [24] rebuilds the volume per frame, it cannot run online; we amortize it"). A delta stated in mechanisms survives review; a delta stated in adjectives does not.

Concurrent work, CVPR norms

With a November deadline and instant preprinting, collisions are routine. The working etiquette: preprints appearing close to or after the deadline are concurrent — cite them once you know of them, in a short "concurrent work" note, without being obligated to outperform them. Two cautions. First, do not use concurrency as a shield for work that predates your submission by months. Second, during the rebuttal you may be pointed at a "missing comparison" that is actually concurrent — answer with dates, calmly, not with indignation.

Double-blind self-citation

CVPR requires citing your own lineage in the third person ("[12] proposed…"), never omitting it — a hole where your prior paper should be identifies you as effectively as a name. Because arXiv posting is legal during review (verified 2026 policy), reviewers may suspect who you are; your job is only to avoid confirming it in the text: no "our previous work [12]", no shared project-page naming, no acknowledgements.

Cite the version of record, and verify the venue

Vision bibliographies rot in a specific way: a paper gets cited as "arXiv 2024" forever after it appeared at CVPR 2025, or gets attributed to CVPR when it was ECCV (NeRF is ECCV 2020; Mask R-CNN and Segment Anything are ICCV; ViT is ICLR). Before the deadline, reconcile the .bib against dblp or the CVF open-access index:

# Bibliography hygiene pass
grep -c '@' refs.bib                                    # how big is the problem
grep -A2 'archivePrefix\|eprint' refs.bib | grep -B2 '20(1|2)[0-9]' | head
# → for each arXiv-only entry older than ~1 year, check dblp for the published venue
grep -in 'cvpr\|iccv\|eccv' refs.bib | awk -F: '{print $1}' | head
# → spot-check 5 random "CVPR" entries on openaccess.thecvf.com before trusting them

Fabricated or wrong citations carry venue-level risk: the verified 2026 author policy makes citations of non-existent material grounds for rejection, without an LLM excuse. Machine-suggested BibTeX gets the same verification as machine-written prose.

A positioning paragraph that does its job

The shape worth copying — question, lanes, deltas, gap — in a fictional example:

How do prior trackers survive long occlusions? Motion-memory methods [3, 17] extend Kalman gates over time, but the gate widens until identity switches dominate beyond ~30 frames. Re-identification banks [8, 24] match appearance on re-entry, yet require high-confidence detections that occlusion exits rarely produce. Query propagation [11] carries identity implicitly but only within the training clip length. None maintains identity through occlusions longer than the model's temporal horizon — the setting we target.

Four sentences, three lanes, each with a mechanism and a failure, ending in the gap the paper fills. A reviewer can disagree with the gap, but cannot say the authors don't know the field — which is the section's actual job.

Sizing and placement

Related work at CVPR is conventionally ~0.5–0.75 of the 8 inclusive pages, either as Section 2 or woven into the intro when space is desperate. Organize by question ("How do prior methods obtain temporal consistency?") rather than by laundry-listed subfield headers; end each paragraph with the sentence that turns the survey into an argument: what remains unsolved that your method solves.

Positioning under fire: rebuttal-time related work

Related work re-opens in January. When a review says "not novel, see [X]":

• If [X] genuinely differs, answer with the mechanism delta in two sentences — never with "we are the first," which invites counterexamples.
• If [X] is concurrent (post-deadline or within the courtesy window), say so with dates and offer a camera-ready citation.
• If [X] actually anticipates you, the honest play is scoping the claim down in the rebuttal rather than defending the indefensible; ACs read the exchange, and a calibrated concession on novelty can coexist with acceptance when the execution and analysis carry weight.
• If [X] does not exist (LLM-flavored reviews hallucinate citations too), report it to the AC via the confidential channel with your search evidence — see cvpr-review-process.

Reverify each cycle

• The current cycle's guidance to reviewers about very recent/concurrent work (the courtesy window is convention, not published law — treat cut-off claims as 待核实).
• Whether citation-format rules changed in the author kit.
• Leaderboard state on your benchmarks in the month before the deadline — SOTA claims age in weeks at this venue.
• The venue's own accepted-paper lists (CVF open access posts them around the conference) for late-breaking nearest neighbors your reviewers will have seen.

Output format

[Coverage] shelves complete: lineage / SOTA / neighbors / cross-venue / concurrent
[Nearest neighbors] <paper → mechanism delta → consequence> ×3
[Concurrent notes] <preprints + dates>
[Self-citation] third-person clean: yes/no
[Bib hygiene] arXiv-only entries checked: <n>; venue misattributions fixed: <n>
[Gaps] <missing comparisons a reviewer will name>

Source: brycewang-stanford/Awesome-Journal-Skills → CVPR-Skills/skills/cvpr-related-work/SKILL.md