acceptodds
Withdrawn from ICLR 2027

ONCE: Patch-Free High-Resolution Depth Refinement via One-Pass Calibration and Enhancement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent high-resolution metric depth refinement methods often recover fine geometric details through patch-based processing, requiring repeated local inference and subsequent fusion. We present ONCE (One-pass Network for Calibration and Enhancement), a patch-free depth booster that converts the output of a frozen monocular metric-depth backbone into a 4K depth map with a single full-frame refinement pass. Unlike patch-based approaches, ONCE decomposes depth refinement into complementary global and local correction stages. It first performs learned global metric calibration to correct image-dependent scale and shift errors in the coarse prediction. It then combines RGB guidance with the coarse depth representation through a dual-encoder refinement architecture and predicts a dense depth-relative correction in log space, enabling spatially varying geometric details to be recovered while preserving the globally calibrated depth structure. This formulation enables native full-frame refinement without explicit patch extraction, repeated overlapping inference, or subsequent patch fusion, while remaining applicable to different frozen metric-depth backbones. On UnrealStereo4K, ONCE-L (ZoeDepth backbone) achieves δ₁ = 99.052%, RMSE = 0.7839, and SEE = 0.6084, establishing state-of-the-art performance among high-resolution depth refinement methods. Compared with the strongest PatchRefiner V2 variant, ONCE-L achieves higher depth accuracy while reducing the number of added parameters from 245.8M in PRV2-C to 192.9M, corresponding to a 21.5% reduction. Consistent improvements across ZoeDepth and DAv2 backbones further demonstrate that ONCE is not tied to a specific coarse depth estimator. Experiments on Cityscapes also show that the same full-frame refinement formulation transfers to real-world driving images. These results demonstrate that global metric calibration, RGB-guided local refinement, and depth-relative correction can be effectively integrated within a single full-frame pass for accurate and efficient high-resolution metric depth estimation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
monocular depth estimation, metric depth estimation, high-resolution depth estimation, depth refinement, depth boosting, patch-free depth refinement

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7JESOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ONCE, the one-pass depth refinement network that combines calibration and enhancement. It shows improved performance over previous patch-based refinement methods on the synthetic UnrealStereo4K dataset.

Critical strengths
  1. The paper is well written and easy to follow.
  2. The proposed one-pass refinement approach offers an interesting alternative to patch-wise refinement methods.
Critical weaknesses
  1. The comparison with existing high-resolution metric depth estimation methods is limited. For example, Depth Pro [1] shows sharp high-resolution metric depth estimation with efficient inference. Since ONCE requires both a coarse depth backbone and an additional refinement network, a comparison in terms of depth accuracy, boundary quality, and end-to-end inference time on real-world datasets can demonstrate practical advantages.

  2. The experimental evaluation mainly relies on the synthetic UnrealStereo4K dataset. As discussed in [1], UnrealStereo4K contains inaccurate depth annotations. If high-resolution depth refinement is also intended for real-world applications, the limited evaluation on real-world datasets makes it difficult to assess the generalization capability of ONCE.

  3. In the real-world Cityscapes experiments, ONCE underperforms PatchRefiner V2 across all reported metrics. Furthermore, the qualitative results in Figure 6 do not clearly show that the improvements in fine-grained depth details observed on UnrealStereo4K are consistently preserved in the real-world setting.

[1] Bochkovskiy, Alexey, et al. "Depth pro: Sharp monocular metric depth in less than a second." International Conference on Learning Representations. Vol. 2025. 2025.

[2] Kwon, Byeongjun, and Munchurl Kim. "One Look is Enough: Seamless Patchwise Refinement for Zero-Shot Monocular Depth Estimation on High-Resolution Images." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

Questions for authors

As mentioned in the weaknesses, could the authors provide a comparison with Depth Pro in terms of depth accuracy, boundary quality, and end-to-end inference time on real-world datasets?

Reviewer_6MUJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses high-resolution metric depth refinement, aiming to improve both metric accuracy and fine-grained geometric details without patch-wise inference. The authors propose ONCE, a single-pass framework consisting of two main components: (1) a Corrector that predicts image-level scale and shift from coarse depth features to calibrate the metric depth, and (2) a Refiner that fuses RGB and depth features to predict bounded log-relative residuals for recovering fine-grained details. Experiments on UnrealStereo4K demonstrate that ONCE outperforms existing patch-based refinement methods in both depth accuracy and boundary quality, while using fewer refinement parameters.

Critical strengths
  • The proposed method achieves superior depth accuracy and boundary quality compared with existing patch-based refinement methods on UnrealStereo4K, while requiring fewer parameters and lower inference time.
  • The experiments on Cityscapes also show promising results, demonstrating consistent improvements over the coarse ZoeDepth baseline and supporting the method's applicability to real-world scenes after domain adaptation.
Critical weaknesses
  • The Corrector estimates global scale and shift solely from coarse depth features, without incorporating RGB information. The rationale for this design is insufficient, and the learned calibration cues may not generalize well to unseen scenes or datasets.
  • The Refiner is essentially a conventional residual prediction network with RGB-depth feature fusion, offering limited research novelty.
  • The authors introduce several predefined bounds for scale, shift, and residual corrections, but do not explain how these values were selected or provide sensitivity analyses.
  • The experiments primarily rely on target-domain training and evaluation. Although results on Cityscapes are reported, the model is also fine-tuned on this dataset. Therefore, its generalization ability to unseen datasets remains unclear.
  • Figures 2 and 3 appear to use rasterized images, causing the text and diagrams to become blurry when enlarged. Vector graphics would improve readability.
  • Both the Corrector and Refiner can potentially compensate for global scale errors, but the paper does not adequately investigate whether the two modules learn distinct and complementary corrections.
Questions for authors

Please see Weaknesses part.

Reviewer_yRe7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ONCE (One-pass Network for Calibration and Enhancement), a patch-free depth booster that refines the output of a monocular metric-depth backbone at 4K resolution in a single full-frame pass. ONCE combines a learned Corrector for image-level metric recalibration with an RGB-guided Refiner for spatial detail recovery. The Corrector predicts scale and shift directly from coarse-depth features and applies them to the coarse metric prediction. The Refiner fuses RGB and depth features through a dual-encoder architecture and predicts bounded log-relative residuals. Experiments show reasonably good improvement over previous SoTA methods.

Critical strengths
  • Experiments show that the effectiveness of the entire system. In Table 1, ONCE achieves the best performance with fewer parameters and faster inference
  • Figure 1 shows the motivation clearly and the visualization comparison in Figure 4 further proved that the claim
  • Using single forward to replace the multiple patch-wise forward is reasonably effective
  • From the system level, it's good to see that the paper combines several simple designs and achieves better performance
  • Overall, it's easy for me to follow the paper
  • It's good to see the authors present the CityScapes results even though it's worse than PRV2
Critical weaknesses
  • Because the patch-free design is an important claim in the paper, I would like to ask the authors if the model can achieve better performance with patch-style inference (stage 1 model before the native-4K fine-tuning) compared with the patch-free one?
  • What is the training cost of the native-4K fine-tuning?
  • As shown in the ablation study table of native-4K fine-tuning, the model without native-4K fine-tuning achieves a much worse performance. It closes the performance gap between the PRV2. As far as I know, PRV2 is trained without this native-4K fine-tuning. It decreases the contribution of other proposed modules. In other words, I have the concern that people can simply achieve performance on par with ONCE by simply tuning PRV2 with native-4K data.
  • There is a lack of more analysis of the Corrector, which is a crucial part of the framework shown in multiple experiments. From the design level, I don't think this module is reasonable. After a metric finetuning on the target dataset, there is no longer a scale and shift gap for the base model's output compared with GT. It's not reasonable and necessary to adopt an another Corrector to estimate a scale and shift to do the rescaling again. However, the experiments show it's important. So I would like to ask more analysis of this output of Corrector. Is the scale always close to 1 and is the shift always close to 0?
Questions for authors

Please reply to my questions in the Critical Weaknesses section.

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

Related papers

Discussion (0)

Sign in to comment.