acceptodds
Withdrawn from ICLR 2027

Single-Image Focal Length Estimators Have a Telephoto Blind Spot, and Their Benchmarks Share It

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Single-image focal length estimators are evaluated almost entirely on wide-to-normal lenses. We audit the four benchmarks recent calibration methods report on: none contains a photograph beyond 243 mm, and on two a constant fitted to the test labels beats every published method. Beyond this range published methods fail. On our benchmark and on three third-party sets that share no images or labels with it, every method trained on rendered panorama crops places at most 2% of telephoto (≥135 mm) photographs within 25% of their true focal length. The failure follows the training range. Fine-tuning only AnyCalib's decoder on photographs spanning 8–1200 mm raises its telephoto accuracy from 0% to 40–48% on all four sets, whereas the same fine-tune on photographs below 110 mm collapses again beyond that limit, so rendering is not the cause. Covering the full range with one model costs accuracy at 20–70 mm. To make the range measurable we release FocalCommons-100K, 100,000 Wikimedia Commons photographs spanning 8–1200 mm, and FocalBench, a 3,177-image benchmark, and we measure how often EXIF focal tags disagree with structure from motion. The label-only statistics behind our audit apply to any regression benchmark and expose such coverage gaps before a model is run.

Primary area:
datasets and benchmarks
Keywords:
focal length estimation, camera calibration, benchmark, dataset, distribution shift, evaluation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5KzxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper points out that existing single-image focal length estimation methods generally fail when applied to long-focal-length scenes, and highlights the limited focal-length coverage of four commonly used calibration benchmarks. The authors introduce FocalCommons-100K and FocalBench, evaluate a variety of existing methods, and investigate the role of training focal-length coverage through controlled fine-tuning experiments.

Critical strengths
  1. The paper addresses an important but insufficiently examined evaluation issue, revealing a systematic telephoto blind spot in existing single-image focal length estimators and their benchmarks.

  2. The paper introduces a large-scale training dataset and a dedicated benchmark to facilitate the evaluation of focal length estimation methods.

Critical weaknesses
  1. Limited novelty. The main contributions are primarily empirical, while the proposed estimators rely on existing architectures and fine-tuning strategies. Although the present paper extends the focal range and provides a substantially more systematic audit, its novelty relative to these closely related contributions (e.g., FocaLens and InFlux++) is not sufficiently established.

  2. The reported performance gains are affected by dataset overlap and label uncertainty. Removing overlapping scenes increases the model's MRE from 19.6% to 27.9%, while most extreme-telephoto labels rely on potentially unreliable EXIF metadata. Moreover, the underlying causes of telephoto generalization failure remain insufficiently explained.

Questions for authors
  1. Could the authors clarify the key contributions beyond existing focal-length datasets and intrinsics-diversity studies?

  2. Given the substantial performance drop after removing overlapping scenes, how can the authors further establish that the reported improvements generalize to genuinely unseen scenes?

  3. How reliable are the extreme-telephoto results given their heavy reliance on EXIF labels?

Reviewer_nCdpOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigate a telephoto blind spot in single image focal length estimation and its evaluation protocols. An audit of four commonly used benchmark splits have limited focal-length coverage and, in two cases return constant targets. The paper also shows that angular field-of-view errors compress focal-length errors at the telephoto end. To broaden evaluation, the authors introduce FocalCommons-100K, a curated photograph collection spanning 8–1200 mm in 35 mm-equivalent focal length, and FocalBench, another 3177-image benchmark.

Critical strengths
  1. The paper identifies an important evaluation blind spot and validates it beyond its own benchmark. The combination of restricted focal coverage, nearly constant targets in some splits, and the compression induced by angular metrics explains why existing evaluation protocols can overlook serious failures. Replication on SR-RAW, FiveK, and InFlux makes the observation substantially more convincing than a result confined to a newly constructed benchmark.

  2. The training interventions provide strong evidence that insufficient focal coverage is an important, repairable cause. Keeping AnyCalib’s backbone frozen and its architecture and loss unchanged, wider range supervision improves telephoto performance across datasets. The range-restricted real-photo control show that replacing rendered images with photographs alone is insufficient. This gives the paper a useful explanatory contribution beyond its dataset contribution.

Critical weaknesses
  1. The long-focal-length labels need stronger validation. Among the 403 FocalBench images at or above 300 mm, 400 use EXIF labels. However, the paper shows that EXIF and SfM focal lengths can disagree, and the cause of this disagreement at long focal lengths remains unclear. This does not invalidate the finding that existing methods fail on telephoto images, but it makes it harder to judge exactly how accurate the proposed models are and how much they outperform other methods.

  2. Much of the proposed model’s advantage comes from training on related scenes. After removing training images from scenes that also appear in the benchmark, the model’s MRE increases from 19.6% to 27.9%. Its advantage over Depth Pro therefore drops from 10.9 to 2.6 percentage points. When uncertainty is estimated by grouping photographs by photographer, this smaller advantage is no longer statistically clear. Although the authors acknowledge this issue, the results provide weaker evidence that the proposed model will consistently outperform existing methods on unfamiliar scenes.

  3. The practical benefit for 3D reconstruction remains untested. The paper shows improved focal length estimate, but doesnot test whether these estimates lead to better 3D reconstructions. A downstream experiment would help establish whether the gains in focal accuracy translate into better recovered geometry, especially for telephoto images.

Questions for authors

1.Have the label at or above 300 mm been checked using a reliable source other than EXIF or SFM? Which conclusions about telephoto accuracy and model ranking can be supported by independent calibrated labels?

  1. Does the telephoto improvement hold across different scenes and videos, rather than being driven by a few groups of similar images? Could you report confidence intervals that treat each scene or video as a group instead of treating every image as independent?

  2. How can we tell whether the mid-range accuracy drop comes from covering a wider focal range, or from choices such as giving more training weight to rare focal ranges? Would it be more accurate to describe this as a trade-off observed under the tested training settings, rather than an unavoidable cost of broader coverage?

  3. Could you compare reconstruction quality using existing methods’ focal estimates and your model’s estimates within the same 3D reconstruction pipeline? Including ground-truth focal lengths as a reference would help show how much the improved estimates benefit reconstruction, particularly at long focal lengths.

Reviewer_dZ3fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

this paper (which has a fairly unusual name) hat four commonly used benchmarks contain no images beyond 243 mm equivalent focal length. Several existing estimators consequently appear successful despite failing on telephoto photographs. The authors introduce FocalCommons-100K and FocalBench, and show that broader training substantially improves telephoto estimation.

Critical strengths

The paper has identified a weakness in common benchmarks for telephoto images. Add more telephoto images significantly improve accuracy. the paper discussed the issues in existing datasets thoroughly, with limitation clearly stated.

Critical weaknesses

The contribution is somewhat weaker for a top calib venue like ICLR. It is well known that even checkerboard calibration can be off when outside its data collection range.

The new dataset construction also raises questions. Of 403 benchmark images at or above 300 mm, 400 rely on EXIF labels. The paper itself finds EXIF–reconstruction disagreement, with an unresolved reversal at long focal lengths. This limits confidence in precise telephoto rankings.

Improvement involves a tradeoff. Broader training worsens AnyCalib’s accuracy at 20–70 mm. The proposed regressor also performs poorly on some standard benchmarks and does not beat Depth Pro overall on FiveK.

Questions for authors

Please see the weakness section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.