acceptodds
Withdrawn from ICLR 2027

Scene Adaptation as Radial Transport in Hyperbolic Space

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Top-leading solutions for scene adaptation typically build on pseudo-label self-training with cross-domain mixing. Though demonstrating steady gains in mean IoU, they remain less reliable on small and rare categories: we observe that self-training improves large categories such as truck and sky while degrading bicycle and rider. In this paper, we reveal the underlying nature of this phenomenon: these approaches regulate how far features move but overlook their direction, and thus fall short in balancing domain alignment against label preservation. To this end, we introduce RadDA, which formulates scene adaptation as radial transport in hyperbolic space. Drawing inspiration from the observation that, in a low-dimensional Poincaré ball, the class of a pixel is encoded in its angle, rare categories occupy the narrowest sectors, and the domain shift is only partially radial, RadDA measures this trade-off by the angular action of the transport. Specifically, RadDA: i) couples source and target pixels via class-aware optimal transport, ii) learns a Riemannian flow under an explicit penalty on , and iii) gates out paths that turn beyond a pixel's angular margin. Since bounds both label corruption and the integration error, one integration step suffices and inference is unchanged. Extensive experiments with a frozen DINOv2 encoder demonstrate that RadDA improves tail-class mIoU on GTA→Cityscapes by 8.0 over the source-only head and 7.0 over self-training at higher mean IoU.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
scene adaptation, unsupervised domain adaptation

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_5pWjOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes RadDA for unsupervised domain-adaptive semantic segmentation, with a particular focus on small and rare classes. It combines a hyperbolic feature representation, class-aware optimal transport, angularly regularized Riemannian flow matching, and a sector gate intended to preserve labels during transport. Under a frozen DINOv2 encoder, the method improves tail-class performance on GTA->Cityscapes. However, the benefits do not consistently extend to other settings, and the theoretical guarantees, necessity of hyperbolic geometry, and distinction from existing methods require stronger justification.

Critical strengths

(1)The paper addresses the important issue that improvements in overall mIoU can conceal poor performance on small and rare categories.

(2)The frozen-encoder protocol helps isolate the effect of feature adaptation. On GTA->Cityscapes, RadDA improves tail-class mIoU by 7.0 points over the reported self-training baseline.

Critical weaknesses

(1)Five reference entries list “VERIFY-AUTHORS” instead of actual author names. These are unresolved placeholders rather than valid bibliographic information. Please verify all affected references and provide complete author lists and persistent identifiers, particularly for ‘Flow matching adaptation for few-shot vision-language alignment’. Their current form makes the related-work discussion and novelty claims difficult to assess.

(2)The main tail-class improvement is demonstrated on GTA->Cityscapes with frozen foundation features. On DarkZurich, tail performance remains below the Euclidean source-only and self-training baselines; on already-adapted hosts, RadDA underperforms self-training. These results substantially limit the scope of the claims about improving small and rare categories.

(3)The manuscript primarily evaluates patch-level predictions and reports full-resolution results for selected configurations, some with only one seed. Please explain the prediction and upsampling procedures and provide consistent full-resolution comparisons across the principal baselines. Since rare classes and small objects are not equivalent, object-size-stratified results would more directly support the small-object claim.

(4)The necessity of hyperbolic geometry is not sufficiently established. Radial–angular decomposition and direction-based classification can also be implemented in Euclidean space. Please compare against a matched Euclidean model using the same class-aware coupling, angular regularization, sector gate, and EMA teacher, to isolate the contribution of negative curvature.

(5) The manuscript is difficult to follow despite the overview in Figure 2. For example, KK appears in the abstract without definition, and the relationships among head training, teacher updates, coupling, flow optimization, and transported-feature supervision are not sufficiently explicit. A concise training algorithm and a clearer statement of the main contributions would improve readability.

Questions for authors

See weakness.

Reviewer_65gSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes RadDA, which fomulates scene adaptations as radial transport in hyperbolic space, with emphasis on improving perfromance on rare classes. The authors argue that existing pseudo-label self-training approaches primarily focus on feature alignment without explicitly controlling the direction of feature movement, and hence fall short in balancing domain alignment against label preservation . RadDA addresses this by representing pixel features in a Poincaré ball, constructing class-aware source-target correspondences through optimal transport, and learning a Riemannian flow regularized by angular action. An angular-margin gate further excludes transport paths that may cross semantic class boundaries.

The authors has done evaluation on GTA→Cityscapes and Cityscapes→DarkZurich using frozen DINOv2 features. The experiments shows comparisons with self-training, Euclidean flow, diffusion, and other adaptation methods. They also incorporated the analysis of angular regularization, integration steps, and computational cost. On GTA→Cityscapes, RadDA reports an improvement of 8.0 points in tail-class mIoU over the Euclidean source-only baseline and 7.0 points over self-training, while also improving overall mIoU.

Critical strengths

The paper points out the important limitation of psudo=label domain adaptaions, wher gains in overall segmentations accuracy may comes at the costof small and rare categories. The prposed fromulation provides an nterresting perspective on balancing domain alignment and semantic preservation. The shown improvements in tail class mIoU are good and the class wise analysis helps in understanding the prectical relevence of the problem. other than this paper also examins how angular regularization effects performance, the integration steps, compare different transport mechanisms under a frozen-encoder protocol.

Critical weaknesses
  1. RadDA shows srong tail class improvments in Table1, GTA→Cityscapes with frozen DINOv2 features, but these improvements do not extend to DarkZurich or the weaker ViT-S encoder. The method also underperforms self training on already adapted DAFormer MIC features. I agree that authors acknowledge these limitaions but it remains unclear what conditions are necessary fro RadDA to be effective.

  2. The papers includes comparison with Euclidean Flow, Diffusion, radial-only transport but the specific benefits of the Riemannian flow-matching formulation remains unclear.

  3. RadDA has achived its best performance with single integration step K, as shown in figure 3, could the authors compare it with a direct learned hyperbolic transformation using the same OT coupling and angular regularization? This would help clarify the additional benefit of the Riemannian flow-matching formulation.

Questions for authors

Please refer to the Critical Weaknesses section.

Reviewer_NszeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RadDA, which formulates scene adaptation as a radial transport problem in hyperbolic space. More specifically, it utilizes class-aware optimal transport to model the relations between source and target pixels, and learns a Riemannian flow in a hyperbolic (Poincaré-ball) feature space. RadDA shows improved performance on scene adaptation tasks like GTA->Cityscapes.

Critical strengths
  1. Using a radial transport problem in hyperbolic space for scene adaptation is interesting.
  2. The results improved on a pretrained DINO backbone.
Critical weaknesses
  1. The problem setting and evaluations are a bit confusing.
  • Table 1 shows that RadDA yields 54.6 mIOU, but it says that Backbone-training pipelines reach 68.3 (DAFormer), 75.9 (MIC), and 79.8 (VFM-UDA++, re-evaluated by us), which are significantly better than ReDA. It is not clear why one should use the proposed ReDA instead of using backbone-training pipelines. Or could ReDA be applied to those backbone-trained models?

  • The gain does not make much sense in realistic settings. On an already self-trained DAFormer, plain self-training beats RadDA by 3.5 mIoU. On MIC, the gain on mean IoU is marginal, and rare-class (tail) mIoU drops by 1.8. There is no tail gain on DarkZurich, and the method fails with ViT-S.

  1. Several citations show "VERIFY-AUTHORS" placeholders.
Questions for authors

The authors are suggested to validate the realistic usefulness of the proposed method by applying and evaluating it on SOTA models.

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

Related papers

Discussion (0)

Sign in to comment.