acceptodds
Withdrawn from ICLR 2027

UniDisp: Learning Unified Hierarchical Stereo Representations for Progressive Disparity Estimation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep stereo disparity estimation has long followed staged pipelines of feature extraction, correspondence modeling, and disparity refinement. Although effective, such pipelines distribute stereo representation learning and disparity inference across multiple modules, while global ambiguous correspondences can remain difficult to resolve from local matching evidence alone. We propose UniDisp, a unified backbone that estimates disparity from stereo representations. This is motivated by a simple hypothesis: disparity can be treated as an intrinsic geometric attribute encoded in stereo representations, rather than a quantity derived from matching within a modular stereo pipeline. We instantiate this idea with a hierarchical stereo encoder that jointly models cross-view displacement cues across spatial scales, a prediction head that maps multi-level stereo representations directly to disparity, organized into a two-stage framework with structurally identical yet functionally complementary encoder–head stages. The first stage performs initial estimation from the raw stereo input, while the second uses implicit warping to construct disparity-aware inputs for residual refinement. This design unifies stereo representation learning, disparity prediction, and refinement into a compact backbone, enabling disparity to be inferred from the learned stereo representations themselves. Extensive experiments demonstrate consistent cross-domain transfer, reaching state-of-the-art performance on several real-world benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
stereo disparity estimation, unified stereo representation, hierarchical representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SzkbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a cost-free stereo disparity estimation method. It adopts the initialization + iterative refinement architecture to regress disparity from coarse to fine. The method mainly follows the WAFT-Stereo. It presents performance on common benchmarks; however, the performance does not reach state of the art performance.

Critical strengths

None

Critical weaknesses
  1. This method is very similar to WAFT-Stereo in the overall architecture, without discussion about its sharp distinction. Moreover, the performance still falls behind WAFT-Stereo considerably in KITTI and cross-domain generation (where WAFT-Stereo is excluded from comparison without any statement).
  2. The presentation does not clearly describe how binocular views are processed by the encoder? They are concatenated along channel dimension, summed, or processed by the encoder independently and then concatenated? This very important architecture design is missing from the paper.
  3. The network presentation gives little information: the HSR encoder is a very common 4-stage backbone design, widely used recent networks like Swin Transformer, ConvNext; the HDP head fuses the 4-stage features of different resolutions like DPT head, although called a different name.
  4. The performance is not good. Cross-domain generalization falls significantly behind SOTA methods like WAFT-Stereo and FoundationStereo. Besides, the comparison with previous methods like BridgeDepth is unfair without indication, which only trained on SceneFlow dataset. Benchmark results on SceneFlow and KITTI are also not satisfying.
Questions for authors
  1. How does the HSR encoder process the binocular views?
  2. What's the core distinction between WAFT-Stereo?
Reviewer_yFGFOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a unified and hierarchical network architecture for stereo matching. The network unifies stereo representation learning, disparity prediction, and refinement into a compact backbone, enabling disparity to be inferred from the learned stereo representations themselves. The network shows good performance on public benchmarks.

Critical strengths
  1. The network architecture is compact and unified.
Critical weaknesses
  1. The idea of hierarchical structure has been used in prior works, such as [1].
  2. I cannot fully understand why the unified representation is better. And there is no experiment to demonstrate it.
  3. On KITTI 2012, the UniDisp seems to be significantly worse than WAFT-Stereo and DispViT+.

[1] Pyramid Stereo Matching Network. CVPR 2018.

Questions for authors
  1. For cross-domain evaluation, what is the performance if UniDisp is trained following prior works using only Scene Flow?
  2. It would be beneficial to compare the inference time with previous methods.
Reviewer_WFTJOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is about stereo vision reconstruction with a neural network. The authors propose UniDisp, a unified backbone that estimates disparity from stereo representations. UniDisdp is composed by a Hierarchical Stereo Representation (HSR) Encoder that directly encodes the joint stereo input through a convolutional hierarchy, and a Hierarchical Disparity Prediction (HDP) Head that decode disparity from the full representation hierarchy and progressively integrates multi-scale geometric cues in a coarse-to-fine manner.

UniDisp is trained from scratch using only synthetic data and then validated on several real-world scenarios.

Critical strengths

The proposed architecture is simple, well-structured, and technically sound.

I think that the main advantage of the proposed method is generalization. Indeed, the experimental results become interesting if one take into account that the network is trained using only synthetic data. This demonstrates strong cross-domain generalization with competitive performance.

Critical weaknesses

Although the proposed approach achieves competitive results, it does not consistently outperform existing state-of-the-art methods like for example on the KITTI benchmark. This may raise concerns regarding the significance of the contribution.

The proposed method predicts dense disparities even in regions where stereo correspondence is fundamentally ambiguous (e.g. texture-less or occluded surfaces). Although the authors outline that long-tail errors remain in challenging regions involving weak textures, they do not investigate whether the resulting predictions correspond to observable geometry or represent hallucinations induced by learned monocular priors.

Questions for authors

Can the author provide a dedicated evaluation of textureless regions, geometric consistency, or predictive uncertainty ?

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

Related papers

Discussion (0)

Sign in to comment.