brycewang-stanford ECCV Experiments
Use this while the experimental plan is still changeable. ECCV's calendar shapes the evidence problem: results freeze in early March, reviews weigh them in May against everything published since, and the field first reads the paper at a September conference — the numbers must still look current six months after the freeze.
The six-month-staleness test
For each headline table, ask: if the strongest lab in this niche publishes their CVPR camera-ready in June, does this table still support the claim in September? Evidence that passes: mechanism-isolating ablations, efficiency frontiers (accuracy vs compute), and generality sweeps across datasets. Evidence that fails: a raw leaderboard number 0.2 points above a moving SOTA. Build the paper's claim on the first kind and let the leaderboard row be corroboration, not the thesis.
Matched-substrate fairness
The first thing a 2026-era vision reviewer checks is whether wins come from the method or from what it was fed:
| Axis to match | Unfair pattern | Fair protocol |
|---|---|---|
| Backbone / pretraining | Your ViT-L vs their ResNet-50 numbers | Re-run the top baselines on your backbone, or add a matched-backbone row |
| Training data | Extra pseudo-labeled or web data only on your side | A same-data row, with the extra-data row labeled as such |
| Input resolution / TTA | Higher test resolution quoted against lower | State resolution and TTA per row |
| Compute / epochs | 4x schedule vs baselines' 1x | Report schedule; add an equal-budget row |
| Foundation-model access | API model in your pipeline, none in baselines | Give baselines the same tool or ablate it out |
One honest matched row protects the paper better than three inflated rows — the mismatched-substrate objection is the most common substantive ECCV review attack and cannot be answered in a one-page rebuttal without a matched number already in hand.
Ablations that isolate, not decorate
- Each claimed component gets exactly one toggle row; combinatorial grids go to the supplement.
- Include the "replace with the dumb version" row (attention → average, learned prior → uniform): it distinguishes mechanism from capacity.
- Ablate on the mid-sized benchmark, not the smallest one, so effects clear
seed noise (
eccv-reproducibilityfor the variance bar).
Qualitative evidence discipline
Vision panels weigh pixels. Ship, in body or supplement: same-scene comparisons against the two strongest baselines; a random-sample grid (not curated) for at least one dataset; and a failure panel tied to the limitations paragraph. A paper with only curated successes reads as hiding something — the failure panel is credibility infrastructure.
Run sequencing toward March 5
T-10 weeks: falsifier first — the experiment most likely to kill the
claim (matched-substrate row on the main benchmark)
T-8: main-table runs launched; seeds x3 on deciding rows
T-6: ablation toggles; efficiency/frontier measurements
T-4: cross-dataset generality; qualitative harvesting begins
T-2: freeze new runs; regenerate all tables from logged results
T-1: random-sample grids, failure panel, supplement tables
T-0 (Mar 5): body tables locked; supplement week polishes, never adds
Launching the falsifier first is the ECCV-specific discipline: with a biennial venue, discovering at T-2 that the matched row erases the win wastes not a cycle but two years.
Output format
[Evidence verdict] mechanism-backed / leaderboard-fragile / incomplete
[Staleness test] <headline table -> survives September? why>
[Substrate audit] <axis -> matched / mismatched -> repair row needed>
[Ablation map] <claimed component -> isolating toggle present?>
[Run queue] <next runs in falsifier-first order with weeks-to-freeze>