brycewang-stanford ECCV Related Work
Use this when auditing novelty positioning for ECCV. The venue's cadence sets the task: since the previous ECCV, two CVPRs and one ICCV have published — for a 2026 submission, that means CVPR 2024, ICCV 2025, and CVPR 2025 (plus early-2026 arXiv) all landed after ECCV 2024. A related-work section that stops at the last even year reads as two years stale.
Coverage sweep, ordered by reviewer expectation
- The direct lineage: the three to five papers your method visibly descends from, with the mechanism-level difference stated per paper.
- The intervening cycles: CVPR/ICCV work from the odd-year gap that attacks the same task — the in-niche reviewer's first check.
- The concurrent band: arXiv papers within a few months of the March deadline; conventions below.
- The classical root: pre-deep or early-deep work the task descends from — European panels notice (and reward) correct genealogy more than most venues.
Concurrent-work conventions around a March deadline
- Treat unpublished arXiv work from the months near the deadline as concurrent: cite it, mark it "(concurrent)", compare qualitatively if feasible, but reviewers cannot require you to beat it. The exact cutoff convention is per-cycle policy — 待核实 in the current author FAQ.
- ECCV 2026's dual-submission rule defined "publication" as peer-reviewed, accepted, and longer than four pages excluding references — arXiv preprints and tech reports did not count as publications, for your eligibility and for what you can be scooped by.
- Your own prior arXiv posting is legal (not an anonymity violation), but the submission must cite it in third person like a stranger's work, and must not cite your own public codebase — that link is treated as an identity leak.
Venue attribution hygiene
ECCV's version of record is Springer LNCS, not CVF open access — and vision bibliographies chronically misattribute across the big three. Check each load-bearing citation against the actual proceedings:
| Error pattern | Example trap | Check against |
|---|---|---|
| ECCV paper cited as CVPR/ICCV | SSD and Perceptual Losses are ECCV 2016, not CVPR | SpringerLink link.springer.com/conference/eccv |
| Sibling paper credited to ECCV | Mask R-CNN is ICCV 2017; ResNet is CVPR 2016 | CVF open access for CVF venues |
| arXiv year vs proceedings year | Preprint 2025, proceedings 2026 | dblp entry |
| "ECCV" for a workshop paper | ECCV workshop ≠ main conference in LNCS subline | The specific LNCS volume front matter |
Making numbered citations argue
LNCS numeric style hides authorship, so the prose must carry the argument:
Weak: "Many works study X [4,9,12,17,21]."
Strong: "Two families address X: geometry-based pipelines [4,9] that
require per-scene optimization, and feed-forward models [12,17]
that trade accuracy for speed; [21] sits between them but
assumes calibrated cameras. We remove that assumption."
Every citation cluster should end with the sentence that says what you do differently — otherwise the cluster is a bibliography, not related work.
Differentiation stress test
Before the freeze, answer in one written sentence each:
- Which single existing paper would a hostile reviewer say already does this?
- What breaks if that paper's method is run on your exact setting?
- Which intervening-cycle paper is closest, and does your experiments
section actually contain it as a baseline (
eccv-experiments)?
If question three's answer is "cited but not compared", fix the experiment or pre-empt in text why comparison is impossible — silence is the reject.
Output format
[Coverage status] <lineage / intervening-cycle / concurrent / classical gaps>
[Attribution audit] <citation -> verified venue, errors found>
[Anonymity conflicts] <own-work citations, repo links to remove>
[Closest rival] <paper -> baseline present? / preempted in text?>
[Cluster rewrites] <bibliography-style cluster -> argument sentence>