acceptodds
Under review as a conference paper at ICLR 2027

GeoMFlow: Geometry-Guided Multi-Frame Optical Flow Estimation with Visual Geometry Priors

Abstract

Modern optical-flow estimators primarily rely on appearance correspondence and task-specific motion representations. As pixel-level motion can be viewed as the image-plane projection of underlying 3D scene motion, geometry foundation models may provide scene-level structural cues that complement these representations. Motivated by this observation, we propose GeoMFlow, a lightweight geometry-conditioning framework that preserves the native matching and cost-volume pathways while injecting spatially aligned geometry features into recurrent flow refinement. Experiments on three standard optical-flow benchmarks demonstrate the effectiveness of visual geometry priors for multi-frame flow estimation. GeoMFlow further achieves an F1-all error of 2.84% and ranks #1 among all optical-flow methods on the KITTI online benchmark.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Optical flow estimation

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_WPwAOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoMFlow, a multi-frame optical-flow estimator that injects pretrained geometry features (from a frozen VGGT encoder) into the recurrent refinement stage via FiLM-style conditioning, while keeping the original matching pathway untouched. It reports state-of-the-art results on Sintel, KITTI, and Spring, including the top rank on the KITTI online benchmark.

Critical strengths
  1. By conditioning only the context branch and leaving the matching pathway and cost volumes completely untouched, the method avoids disturbing the pretrained correspondence representation, and the plug-in stays lightweight and architecture-agnostic. This makes it easy to attach to modern flow estimators with minimal risk, which is a practical merit of the design.

  2. Comparing DINOv2, DAv2, and VGGT under an identical backbone and training protocol isolates the effect of the representation source itself, and the same conditioning interface is shown to transfer between a multi-frame and a two-frame architecture.

Critical weaknesses
  1. The method is essentially a combination of existing components, that is, VGGT features, align them, and inject them into the context branch with an off-the-shelf FiLM layer. Similar ideas have already been explored by concurrent works like FlowSeek(depth information), so the novelty is not that strong.

  2. It's not clear to me why,why,why geometry priors actually help. The paper only argues that flow is the image-plane projection of 3D motion, so geometry features should help, but what specific information do VGGT features actually provide for flow?

There are no probing or visualization experiments, and I honestly don't see the mechanism.

Actually, it's similar to the first question. Why use geometry, model is work?

  1. In Table 5, the w/o Geo baseline is 0.381 EPE; with geometry added, Full FT only reaches 0.377, while Frz Both jumps to 0.329. This large gain seems to come from the training recipe of "freezing encoders and only training the adaptation ++ refinement modules", not from geometry conditioning itself.

However, the paper never includes a w/o Geo + Frz Both control, so this explanation cannot be ruled out.

Questions for authors

Please answer the question carefully and convince me why geometry is effective?

After reading the entire text, the most memorable thing for me is that geometry is easy to use, but I don't know why or under what circumstances it is easy to use

Reviewer_bq9BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents GeoMFlow, a geometry-conditioning framework that improves existing optical flow methods. It uses features from a frozen VGGT encoder to condition recurrent context features through a spatially aligned FiLM-style adapter, while preserving the pretrained correspondence representation. Applied to MEMFOF, GeoMFlow achieves the state-of-the-art KITTI result and top Spring result.

Critical strengths

The performance on KITTI is strong. On KITTI, GeoMFlow achieved state-of-the-art results on both EPE and Fl-all.

Critical weaknesses
  1. GeoMFlow does not consistently achieve state-of-the-art performance across the Sintel metrics. On Spring, its gains over MEMFOF are mixed: two metrics improve, one remains unchanged, and one metric is worse. Moreover, the Spring comparison omits several relevant baselines such as WAFT, which have very strong performance on Spring.

  2. The claim that geometry conditioning is “lightweight” requires clearer justification. Although the conditioning adapter itself may be small, GeoMFlow also requires VGGT feature extraction, whose computational overhead should be included in the full system’s inference cost. The paper does not explicitly clarify whether the reported runtime includes VGGT feature extraction. A detailed breakdown of runtime and memory usage is needed to support the “lightweight” claim.

Questions for authors
  1. I would need clarifications for the concerns I raised in the Critical Weakness section.
  2. Why the geometry conditioning is only applied to context encoding? Would it be helpful to also introduce that into the cost volumes?
Reviewer_WrejOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a geometry-guided optical flow estimation framework. The proposed method leverages a geometry foundation model to modulate geometry and context features using FiLM. The modulated features are then injected into the cost volume for flow refinement.

Critical strengths
  • Lightweight architecture and strong performance. The proposed framework is lightweight, and extensive benchmark evaluations demonstrate its competitive performance across various datasets.
  • Strong generalization capability. The zero-shot evaluations demonstrate the generalization capability of the proposed method, highlighting its potential applicability to unseen datasets.
Critical weaknesses
  • Limited technical novelty. Despite its strong performance, the technical novelty of the proposed method appears limited. The performance improvements seem to be primarily achieved by combining existing foundation models rather than introducing fundamentally new techniques. While achieving strong performance is important, I believe the paper should provide deeper technical insights to make a more significant contribution to the ICLR community.
  • Unclear effectiveness of the proposed training strategy. According to Table 5, freezing all encoders achieves the best performance. This raises concerns about whether training the encoders using the proposed strategy actually degrades performance. The authors should provide further analysis to clarify this observation and justify their training strategy.
  • Insufficient analysis of geometry guidance. Table 4 shows that using models such as DINOv2 for geometry guidance results in worse performance than using no geometry guidance. This observation requires further analysis to explain why certain pretrained representations negatively affect optical flow estimation and what properties make a foundation model suitable for geometry guidance.
  • Lack of clarity in mathematical formulations. The equations are not numbered, making the paper difficult to follow and reference. In addition, the variables UU and cc in the equation on line 238 are not clearly defined. The authors should provide explicit definitions for all variables and number the equations to improve readability.
Questions for authors

Please refer to the concerns raised in the Weaknesses section, particularly regarding the limited technical novelty, the effectiveness of the proposed training strategy, and the analysis of geometry guidance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.