ECCV Reproducibility
Use this before the ECCV paper freeze. ECCV publishes through Springer LNCS
with no standing mandatory reproducibility checklist across cycles (whether
the current cycle adds one: 待核实 against the live author guidelines), so
the reproducibility bar is enforced socially: by reviewers who try to match
your numbers, and by the two-year gap before you could publish a correction
at the same venue.
The two-year checkability horizon
A CVPR paper's errors are challenged within a year; an ECCV paper sits as
the venue's latest word on the topic until the next even year. Write the
paper so a lab starting from only the PDF plus supplement in 2027 can
rebuild the result — that is the horizon reviewers implicitly price in.
Recipe ledger (goes in paper or supplement, never nowhere)
| Ingredient |
Minimum disclosure |
Common ECCV-draft omission |
| Training schedule |
Optimizer, LR schedule, epochs/iterations, batch size, augmentations |
Augmentation list "standard" with no definition |
| Initialization |
Pretrained checkpoint identity + source |
"ImageNet-pretrained" without which checkpoint |
| Data |
Dataset version, split definition, filtering rules |
Custom val split described only as "held out" |
| Evaluation |
Metric implementation source, input resolution, TTA on/off |
Resolution mismatch between method and baselines |
| Compute |
GPU type, count, wall-clock, total runs behind the paper |
Only the final run's cost reported |
Foundation-model era pinning
Modern ECCV pipelines sit on moving substrates. Pin all of them by exact
identity, because "CLIP features" is not reproducible information:
# pinned-substrate block for the supplement
backbone: dinov2-vitl14, weights sha256:<hash>, source: <url>
vlm: <model name + exact release tag>, accessed 2026-02
sam_variant: <checkpoint id>
inference: fp16, single-crop, resolution 518x518
api_models: none # if any API model is used, record date + version string
An API-served model that silently updates invalidates comparisons; record
access dates and version strings, and prefer frozen open-weight substrates
for headline tables.
Variance honesty on benchmark deltas
- A +0.3 mAP or +0.2 mIoU headline delta needs seed evidence: report
mean ± std over ≥3 seeds for your method and your strongest baseline,
or scope the claim down.
- State which numbers are your re-runs versus quoted from prior papers —
mixed provenance inside one table is a classic silent irreproducibility.
- If full re-training is too expensive to repeat, say so and report seeds
on the cheapest deciding component (e.g., the head, not the backbone).
Split the story across the 14 pages and the supplement
- Body: enough recipe to judge plausibility — schedule summary, data
versions, compute order-of-magnitude.
- Supplement: the full ledger, per-experiment configs, the pinned-substrate
block, and negative-result notes ("we tried X at lr=1e-3, diverged").
- Code archive: configs as files, not prose; the paper should never be the
only serialization of a hyperparameter.
Honest-failure statement
One paragraph reviewers reward at this venue: name the regime where the
method breaks (small objects, low light, out-of-distribution categories),
with a pointer to a supplement figure showing it. It signals the numbers
were probed rather than curated.
Output format
[Repro grade] rebuildable-from-paper / rebuildable-with-code / not-rebuildable
[Ledger gaps] <schedule / init / data / eval / compute rows missing>
[Substrate pinning] <unpinned dependency -> exact identity to record>
[Variance status] <headline delta -> seed evidence present?>
[Placement plan] <body vs supplement vs code archive>
1---2name: eccv-reproducibility3description: Use when hardening the reproducibility story of an ECCV paper — training recipes and schedules readers can re-run, dataset versioning and split provenance, pinned foundation-model dependencies, compute disclosure, and seed/variance honesty for benchmark deltas, sized for the 14-page LNCS body plus supplement.4---56# ECCV Reproducibility78Use this before the ECCV paper freeze. ECCV publishes through Springer LNCS9with no standing mandatory reproducibility checklist across cycles (whether10the current cycle adds one: 待核实 against the live author guidelines), so11the reproducibility bar is enforced socially: by reviewers who try to match12your numbers, and by the two-year gap before you could publish a correction13at the same venue.1415## The two-year checkability horizon1617A CVPR paper's errors are challenged within a year; an ECCV paper sits as18the venue's latest word on the topic until the *next even year*. Write the19paper so a lab starting from only the PDF plus supplement in 2027 can20rebuild the result — that is the horizon reviewers implicitly price in.2122## Recipe ledger (goes in paper or supplement, never nowhere)2324| Ingredient | Minimum disclosure | Common ECCV-draft omission |25|---|---|---|26| Training schedule | Optimizer, LR schedule, epochs/iterations, batch size, augmentations | Augmentation list "standard" with no definition |27| Initialization | Pretrained checkpoint identity + source | "ImageNet-pretrained" without which checkpoint |28| Data | Dataset version, split definition, filtering rules | Custom val split described only as "held out" |29| Evaluation | Metric implementation source, input resolution, TTA on/off | Resolution mismatch between method and baselines |30| Compute | GPU type, count, wall-clock, total runs behind the paper | Only the final run's cost reported |3132## Foundation-model era pinning3334Modern ECCV pipelines sit on moving substrates. Pin all of them by exact35identity, because "CLIP features" is not reproducible information:3637```yaml38# pinned-substrate block for the supplement39backbone: dinov2-vitl14, weights sha256:<hash>, source: <url>40vlm: <model name + exact release tag>, accessed 2026-0241sam_variant: <checkpoint id>42inference: fp16, single-crop, resolution 518x51843api_models: none # if any API model is used, record date + version string44```4546An API-served model that silently updates invalidates comparisons; record47access dates and version strings, and prefer frozen open-weight substrates48for headline tables.4950## Variance honesty on benchmark deltas5152- A +0.3 mAP or +0.2 mIoU headline delta needs seed evidence: report53 mean ± std over ≥3 seeds for your method *and* your strongest baseline,54 or scope the claim down.55- State which numbers are your re-runs versus quoted from prior papers —56 mixed provenance inside one table is a classic silent irreproducibility.57- If full re-training is too expensive to repeat, say so and report seeds58 on the cheapest deciding component (e.g., the head, not the backbone).5960## Split the story across the 14 pages and the supplement6162- Body: enough recipe to judge plausibility — schedule summary, data63 versions, compute order-of-magnitude.64- Supplement: the full ledger, per-experiment configs, the pinned-substrate65 block, and negative-result notes ("we tried X at lr=1e-3, diverged").66- Code archive: configs as files, not prose; the paper should never be the67 only serialization of a hyperparameter.6869## Honest-failure statement7071One paragraph reviewers reward at this venue: name the regime where the72method breaks (small objects, low light, out-of-distribution categories),73with a pointer to a supplement figure showing it. It signals the numbers74were probed rather than curated.7576## Output format7778```text79[Repro grade] rebuildable-from-paper / rebuildable-with-code / not-rebuildable80[Ledger gaps] <schedule / init / data / eval / compute rows missing>81[Substrate pinning] <unpinned dependency -> exact identity to record>82[Variance status] <headline delta -> seed evidence present?>83[Placement plan] <body vs supplement vs code archive>84```
brycewang-stanford