"Have You Considered Collaborating With a University Academic?"

The author may well benefit from collaborating with a university academic to help with the experiments, use of language and structure of this paper.

— Reviewer #4, IVCNZ

That sentence is the reason this post exists, and the reason it names the conference. Most of what I’ve written about publishing so far has kept venues anonymous, because the point was the mechanism rather than the venue. This time the mechanism and the venue are the same thing: a single-blind review process at IVCNZ (Image and Vision Computing New Zealand) let a reviewer recommend a change of author as the remedy for a paper, and nothing in that process stopped it.

This is my account, from my side of the submission, and I’m not going to pretend to be neutral about it. I was offended by that sentence, and I was shocked by how little the rest of the process held up. But I’ll go through all three reviews properly, because they are not equally bad, and pretending they are would make the one that matters easier to dismiss.

The paper, briefly

The paper is Independent Samples, Correlated Variance: A Learnable Cross-View Cue in Path-Traced Stereo Data. Synthetic stereo datasets are increasingly rendered with physically based path tracers, and the pipelines that consume them assume the Monte Carlo noise in the two views is independent. At the level of individual samples it is. But the per-pixel variance field — how hard each pixel is to integrate — is strongly correlated across the two views once you warp by ground-truth disparity (ρ ≈ 0.75 against 0.36 unwarped, replicated on a second renderer), and a small siamese probe trained on nothing but those fields learns to match with it.

It is a characterisation paper. It measures a property of the data, explains where it comes from, and shows a network can use it. It does not propose a stereo method, and it says so.

The review was single-blind. My name and “Independent Researcher” were on the first page.

Reviewer #5: lost, and partly my fault

I’ll start with the review I take seriously.

I think there is an interesting idea in here, but hte presentation of the paper is so dense that it is hard for the reader to determine. If the paper is discussing what I think it is, then it is certainly likely that path-tracers’ local nosie properties depend on the surface being rendered, and that this may be a signal that could be used to estimate disparity. The causes of this signal are not, however, clearly explored.

The density complaint is fair, and I own it. My abstract carries more than a dozen numbers. The first page alone introduces 0.754 ± 0.016, −0.0004, 0.38, 78.8% and 2.40 percentage points before the reader has been told why any of them matter. I wrote it for a reader who already believed the question was worth asking, and a reviewer coming in cold had to carry all of that before getting to the part that explains it.

But the substantive criticism — that the causes of the signal are not explored — is not right, and it’s worth being precise about why, because the way it’s wrong is itself evidence of the density problem.

Section VII is the causal account. It splits the radiance integrand at a surface point into a view-independent part (shadowing, indirect illumination, caustics on diffuse surfaces) and a view-dependent part (Fresnel, glossy and refractive terms). After warping, both pixels sample the same surface point, so the two variance fields agree to the extent the view-independent part dominates. That makes a prediction you would not guess in advance: glass, which has roughly seven times the variance of everything else, should show weaker cross-view structure, not stronger. It does, and the same ordering shows up at four independent levels of evidence — the correlation on Mitsuba, the correlation on Cycles, the learned probe on variance fields, and the learned probe on single-render residuals.

The reviewer’s own guess — “local noise properties depend on the surface being rendered” — is the first half of that section. The paper goes further and says which part of the surface response carries the structure, and checks it.

So I don’t think the reviewer made a careless mistake. I think they got lost before page five, which is exactly what they said happened. The explanation was there; I made it too expensive to reach. That’s a writing failure, and it’s mine. The “Poor” on technical soundness I read the same way: no technical defect is named, and I think it follows from being lost rather than from finding one.

This is what a review that engages looks like, even when it’s wrong. It summarises the paper in its own words, it says “I may be mistaken”, it states a hypothesis about the content, and it gives me something I can actually fix.

Reviewer #3: a review that wasn’t, counted anyway

Summary of Contributions: Not able to comment. Expertise: Passing Knowledge Overall Rating: Possibly Accept Review: Not able to review; too far out of my field.

I have no complaint against this reviewer. Saying “this is outside my field” is the honest thing to do.

My complaint is that it reached me as a review. The form presumably forces a rating in every field before it can be submitted, so the reviewer filled them in — Fair, Fair, Fair, Possibly Accept — and those numbers are noise generated by a form, not a judgement about the paper. Someone running the process saw “Not able to review” and let it stand as one of three reviews instead of replacing it.

The numbering makes this harder to excuse. The reviews I received are #3, #4 and #5. I don’t know what happened to #1 and #2, but the numbering implies the paper was assigned to at least five people. This was not a venue that had run out of reviewers. Effectively, the paper was evaluated by two people, and one of them is the subject of the next section.

Reviewer #4: line by line

Here is the entire review text.

This paper presents an interesting discussion about learnable cross-view cue in path-traced stereo data – but would benefit from clear simple direct quantitative comparisons with the results of prior research – and stated in both the abstract and conclusion. The author may well benefit from collaborating with a university academic to help with the experiments, use of language and structure of this paper.

Expertise: Expert. Overall rating: Reject.

Three sentences. Let me take them in order.

“An interesting discussion.” This restates the title. The one word it adds is discussion, which is a strange word for a paper whose body is scene-level bootstrap intervals, a seed-count sweep that bounds the reported value from below, a placebo control, a difference-in-differences design with a sign test, and a paired replication on a second renderer. It’s the word you’d use if you’d read the abstract and the conclusion.

“Clear simple direct quantitative comparisons with the results of prior research.” I want to be fair here, because there is a reasonable request inside this sentence. The paper explicitly does not measure whether relying on this cue helps or hurts a deployed stereo network on real data, and a vision reviewer can reasonably ask: then why should I care? The conclusion also contains no numbers at all. If the sentence meant “show me downstream impact, and put it where I’ll see it”, that’s a legitimate thing to want, and a better version of the paper might address it.

But that isn’t what it says, and the literal request has no object. The related-work section argues, across four neighbouring literatures, that nobody has measured this quantity — that is the point of a characterisation paper. There is no prior ρ to put next to mine. A reviewer who rates themselves Expert and thinks otherwise could have named the paper I should compare against. They didn’t name one, or one number, or one section.

“The author may well benefit from collaborating with a university academic to help with the experiments, use of language and structure of this paper.”

Here is a test anyone can run in their head. Take the same PDF and change one line on the first page, from “Independent Researcher” to “University of Somewhere”. Does that sentence still get written?

It can’t. There’s nothing for it to attach to. A reviewer with a problem with the experiments names the experiment. A reviewer with a problem with the language quotes the sentence. A reviewer with a problem with the structure — which is a fair problem, see Reviewer #5 — says where it breaks. What this sentence does instead is prescribe a different kind of author. The only input it requires is the affiliation line, and in a single-blind review the affiliation line is exactly what the reviewer can see.

Of the three things it lists, experiments is the one I’d least expect anyone to pick. The experiments are the most defended part of the paper. Every interval is a scene-level bootstrap, never a pixel count. Every learnability result is paired with a decorrelation control and a placebo. The cross-renderer test checks that the two scene ports agree on material masks and disparity before it claims anything. I’m not saying the experiments are beyond criticism. I’m saying the review doesn’t make one.

I’ll say this plainly: I was offended. Not by the Reject. I’ve had rejections I agreed with and rejections I didn’t, and neither felt like this. What offended me was being told, in a document that is supposed to be about my paper, that the fix was to go and find someone with a university affiliation to supervise my experiments, my English and my structure. “May well benefit from” is a polite construction. It read to me as contempt.

I can’t see into the reviewer’s head, and I don’t need to. That is the part that should worry people running review processes: whether it was meant as a sneer or as sincere advice, the sentence does the same thing. It substitutes the author’s affiliation for an argument about the paper.

The scores make this visible. Here is what each rating is supported by in the text of the review:

The reproducibility score is worth a moment, because it’s the one place where a “Poor” could in principle be fair without being explained. It isn’t. IVCNZ’s submission form had no field for code or supplementary material, so there was nowhere to put a repository even if I’d wanted to. What the paper can carry, it does: renderer versions, resolution, baseline, field of view, path depth, sample budget, seed count, the exact warp convention, the masking rule, the probe architecture and parameter count, the training budget and the train/test split. And an earlier version of this paper had already been through three full reviews at another venue, as the next section describes. None of those reviewers raised reproducibility.

The same work, double-blind

Before IVCNZ, an earlier version of this paper went to BMVC 2026, which reviews double-blind on OpenReview. I should say clearly how that went, because this is not a story where a good venue accepted the paper and a bad one didn’t. BMVC rejected it too.

But the rejection looked completely different. Three reviewers, a rebuttal round, and a meta-review from an area chair. The reviewers converged on the same handful of problems, and each one was specific:

• No trained network was shown to exploit the cue, so the “shortcut” framing was only suggestive. All three raised this. One asked me to “compare training with normal versus decorrelated synthetic data.”
• Everything was on one renderer, Mitsuba 3, with the seed count fixed at N = 30 and never swept.
• The claim that real sensors lack the structure was asserted, not measured.
• The writing was hard to follow. One reviewer said exactly how: the abstract “front-loads numbers” before the problem, the gap and the insight.

One reviewer listed what would change their score. That is a review I can work with, and I did. The IVCNZ version is the result: a second renderer with the scenes ported item by item, a seed-count sweep with an attenuation fit that shows the reported value is a lower bound, a learned probe trained on normal versus decorrelated data — the experiment that reviewer suggested — with a placebo control, a difference-in-differences design under the realistic single-render condition, and a sharper argument from the sensor noise model in place of the bare assertion. I didn’t do everything. Training a full, named stereo network on two path-traced corpora at scale is a dataset-construction project, and the paper says so in its limitations. But most of what they asked for is there.

Now set the two sets of reviews side by side.

One BMVC reviewer rated themselves Expert, the same rating Reviewer #4 gave themselves. That reviewer wrote five numbered weaknesses, argued with Eq. 4 directly, and also wrote that the paper provides “rigorous theoretical background to support their claims”. Another called the math “simple and correct”, said the intervention was “well designed”, and said “the epistemics are careful”. These are people who recommended rejecting the paper. Not one of them suggested I find a co-author. Not one mentioned an affiliation, because there wasn’t one for them to see.

Read next to them, Reviewer #4’s second sentence makes more sense, and less. More, because “comparisons with the results of prior research” is probably reaching for what all three BMVC reviewers asked for: show the effect on a real, named stereo network. Less, because the BMVC reviewers could say which experiment, on what, and why it mattered, and Reviewer #4 couldn’t.

I also have to own something the BMVC reviews make obvious. The density that lost Reviewer #5 is not a new problem. All three BMVC reviewers flagged the writing, and in my rebuttal I promised to put the problem and the insight before the numbers. I fixed the order. I didn’t fix the load. Reviewer #5 got lost in the same place the others did, and that part of this story is on me.

This is one paper and two venues, not a controlled study, and the IVCNZ version is a different paper from the BMVC one. But the comparison does isolate one thing. The same research, rejected twice, got substantive criticism when the reviewers couldn’t see who I was, and a recommendation to find a university academic when they could.

What single-blind costs an independent researcher

The usual argument for single-blind review is that reviewers can’t really be blinded anyway, and knowing who the authors are helps them calibrate. Calibrate is the polite word. In practice, the affiliation line becomes a prior, and for most authors that prior is neutral or positive, so nobody notices it’s there.

For someone with no institution, the prior is the review. There’s no lab name to lend credibility, no advisor who presumably checked the work. There’s a line that says Independent Researcher, and a reviewer who has not engaged with the content has one obvious thing left to say.

I wrote earlier about how a venue’s published scope fails to bind what its editors actually do. This is the same gap at a different layer. Every conference’s reviewer guidelines say some version of “evaluate the work, not the authors”. None of that is enforced at the only point where it could be: an area chair reading the reviews before they go out, seeing a recommendation that the author find a different kind of co-author, and sending it back.

Nobody did that here. At BMVC an area chair read three reviews and a rebuttal and wrote a paragraph explaining the decision. At IVCNZ there was no meta-review. No area chair, no programme chair, nobody wrote a single line summarising the reviews, weighing them, or noting that one of the three had declined to review at all. The reviews reached me exactly as written: a non-review counted as a vote, a review whose only concrete recommendation was about the author, and one honest reader who got lost. That was the whole decision.

What I’m taking from it

I’ll be honest about the part that surprised me most. I went in expecting a regional conference to be smaller and less competitive than the big venues, and I was fine with that. I didn’t expect the floor to be this low. A review that admits it can’t review is passed through with a score. A self-declared expert submits three sentences, none of them technical, and recommends rejection. Nobody reads any of it before it goes out. That is not a strict venue or a harsh one. It is a process that isn’t being run.

I won’t submit to IVCNZ again. Not because a paper got rejected — papers get rejected, and this one may well need another round — but because the process showed me what it will let through, and I have no reason to think it would be different next time.

From Reviewer #5 I’m taking the actual work: the paper needs to let a cold reader reach Section VII without carrying fifteen numbers. That’s a real revision, and it will make the paper better.

From Reviewer #4 I’m taking nothing about the paper, because there was nothing about the paper to take. I’m keeping the sentence, though. It’s at the top of this post.