acceptodds
Under review as a conference paper at ICLR 2027

Causal-Flow: Deconfounding Dark Optical Flow Estimation

Abstract

Dark Optical flow estimation is fundamentally challenged by the entanglement of true motion and degradation-induced confounders in the observed information. Existing paradigms treat this entangled representation holistically, applying enhancement or suppression without explicitly identifying the degradation contribution, thereby incurring an inherent trade-off between preserving true motion and suppressing spurious responses. In this paper, we propose Causal-Flow, a causal decoding framework built on a dual causal design. First, a Causal Intervention Module (CIM) performs front-door adjustment, projecting motion features onto a low-rank subspace spanned by true motion patterns to remove high-dimensional stochastic confounders. Second, a Counterfactual Gating Unit (CGU) estimates degradation-induced spurious motion and derives adaptive spatial and channel gates from causal residuals to enhance true motion and suppress ambiguities. By embedding the CGU into each gate of the GRU-based decoder, Causal-Flow enables single-step decoding without iterative refinement, circumventing error accumulation and achieving state-of-the-art performance on dark-scene optical flow benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dark Optical Flow Estimation, Causal Reasoning

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_kY2SOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses optical flow estimation under low illumination, where sensor noise, weak textures, and other image degradations can create unreliable correspondence evidence. The proposed Causal-Flow architecture introduces two modules into a flow-matching-based optical flow decoder: 1) Causal Intervention Module (CIM) which constructs a mediator feature, applies learned spatial weighting, and projects features onto a learned low-rank channel subspace; 2) Counterfactual Gating Unit (CGU) which predicts a degradation baseline from motion and context features, subtracts it to obtain a residual, and uses this residual to generate spatial and channel gates. The proposed method has been evaluated on FCDN and VBOF under two training settings, with additional qualitative comparisons on four real-world low-light datasets. Experiment shows that compared with the strongest reported baseline FlowFM, proposed Causal-Flow reduces FCDN EPE from 0.87 to 0.81 when trained on FCDN, and from 1.06 to 1.02 when trained on Mix.

Critical strengths
  1. Consistent benchmark improvements: the proposed method improves over existing methods consistently, as shown in table 1 and table 2. Furthermore, the cross-dataset setting: training on FCDN and evaluating on VBOF, is particularly relevant to robustness against changes in camera characteristics and degradation patterns.
  2. The ablation study is clear and extensive: table 3 examines the effects of removing entire modules (CIM/CGU/CRD), spatial weighting, the low-rank basis, individual gates, and initialization choices, which clearly demonstrate the effectiveness of each module under different settings.
Critical weaknesses
  1. The 'Causal reasoning' in Figure 2 is somewhat confusing, the claim that "Our modules block the confounding path and estimate P(Y |do(X)), so that the flow is determined only by the true motion" is not clearly illustrated in Figure 2. Furthermore, in Sec. 3.2, the paper introduces a mediator and invokes front-door adjustment, but defining an intermediate neural feature does not, by itself, establish that it is a valid causal mediator. Basically, it is a reasonable motivation for separating useful motion information from nuisance responses, but attribute it to causal reasoning may need more justification
  2. The argument that true motion is low-dimensional while degradation is high-dimensional is plausible for some noise models, but is not universally valid. Low-light degradation can contain structured components such as banding, fixed-pattern noise, flare, and correlated artifacts. Conversely, valid motion can contain discontinuities, small objects, thin structures, and complex local displacements.
Questions for authors
  1. What is the full derivation of Equation 2, including assumptions on the prediction function?
  2. In Sec.3.2 the rank is set at half the channel count, how sensitive are results to this choice?
Reviewer_QpRSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Causal-Flow for optical flow estimation under low-light conditions. Building on FlowFM, it introduces a Causal Intervention Module (CIM), which combines query-guided weighting with low-rank projection, and a Counterfactual Gating Unit (CGU), which modulates motion features using a learned residual. These components are integrated into a GRU-based decoder with single-step inference. Experiments show improvements on FCDN and VBOF, supported by component ablations and qualitative results on real-world scenes.

Critical strengths
  1. Distinguishing genuine motion evidence from degradation-induced responses addresses an important challenge in low-light optical flow estimation.
  2. The method improves over FlowFM under both training settings, reducing VBOF EPE from 13.28 to 12.96 when trained on FCDN and from 5.24 to 4.94 with mixed training.
  3. The ablations examine multiple elements of CIM and CGU, while the qualitative comparisons illustrate improvements around motion boundaries and degraded regions.
Critical weaknesses
  1. The paper does not establish the required mediation conditions; in particular, Eq. (5) preserves a direct input residual that bypasses the proposed mediator. It also does not demonstrate why query-weighted projections estimate the required marginal expectations, or why the NWGM approximation in Eq. (2) applies to this nonlinear flow-regression decoder.
  2. The paper assumes that true motion occupies a low-dimensional subspace while degradation lies outside it, without analyzing the learned representations. Comparisons with ordinary low-rank adapters and spatial/channel attention under matched parameter budgets, together with controlled degradation experiments, would help distinguish the proposed causal explanation from conventional feature filtering.
  3. Real-world evaluation is qualitative and lacks flow ground truth, so sharper silhouettes alone cannot establish greater motion accuracy. The benchmark gains are consistent but sometimes small, and no variability across training runs is reported.
  4. The manuscript does not fully specify the noisy-flow construction, initialization of f_0, or whether baseline training settings are matched. Additionally, increasing decoder iterations from one to two changes the parameter count in Table 3, making it difficult to attribute the degradation specifically to iterative error accumulation.
Questions for authors

Please refer to the concerns raised in the Weaknesses section.

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

This paper studies optical flow estimation in dark scenes by describing the observed motion evidence as a combination of true motion and degradation-induced interference. Building on FlowFM, the authors introduce a Causal Intervention Module (CIM), which uses query-guided low-rank projection to filter noisy features, and a Counterfactual Gating Unit (CGU), which predicts a degradation baseline and modulates the gates of a GRU-based decoder. The resulting model performs a single decoding update and retains relatively low inference cost. Experiments on FCDN, VBOF, and several real-world low-light datasets report consistent improvements over existing methods.

Critical strengths
  1. Dark-scene optical flow is practically important and technically challenging, and the paper presents the motivation clearly.
  2. The method is built on a strong recent baseline, FlowFM, while keeping relatively low computational cost: 35 ms and 21.7 GFLOPs according to Table 4.
  3. The evaluation covers FCDN-only and mixed-dataset training, cross-dataset generalization, individual camera subsets, and qualitative real-world examples.
Critical weaknesses

Weaknesses and Questions

  1. The paper claims to implement front-door adjustment through a learned mediator feature, but it does not establish the necessary causal identification conditions. In particular, it is unclear whether the mediator blocks all directed paths from the observed motion features to the estimated flow. Furthermore, Eq. (2) approximates the standard front-door marginalization by applying the prediction function to expected features, without sufficiently explaining the assumptions underlying this approximation. The expectations in Eqs. (3)–(4) are then computed using attention-weighted low-rank projections rather than explicit marginalization over the input distribution. These concerns raise questions about whether CIM truly performs causal intervention or primarily acts as a conventional feature refinement module.

  2. The predicted degradation baseline (f_d) is learned solely through the final optical flow loss, without counterfactual data or independent supervision. Therefore, it is unclear whether (f_d) captures only degradation-induced responses or also contains true motion information. Moreover, although Eq. (6) computes the residual (f_m^{CIM}-f_d), Eq. (8) uses this residual only to derive gating weights, while the original purified features are used for the final update. This differs from the claimed explicit removal of degradation components. Further analysis is needed to justify the counterfactual interpretation of CGU.

  3. While the ablation studies demonstrate the effectiveness of CIM and CGU, they do not directly verify whether the modules successfully separate true motion from degradation-induced interference. Controlled experiments that vary degradation levels while keeping the underlying motion unchanged would provide more convincing evidence for the claimed causal disentanglement.

  4. The improvements may arise from low-rank feature regularization and spatial-channel attention rather than causal reasoning. Although the authors compare CIM and CGU with alternative modules, further comparisons with conventional denoising, low-rank projection, and attention mechanisms under comparable model capacities would help clarify the actual contribution of the proposed causal design.

  5. The evaluation on real-world datasets relies primarily on qualitative comparisons without optical flow ground truth. Additional objective evaluations, such as forward-backward flow consistency or occlusion-aware warping errors, would provide stronger evidence for the claimed real-world generalization.

  6. The experimental evaluation focuses primarily on optical flow accuracy for individual frame pairs, without assessing temporal consistency across consecutive frames. Furthermore, no video demonstrations are provided to illustrate the temporal stability of the estimated flow. Additional temporal consistency analysis or video results would help establish the method's reliability in real-world dynamic scenes.

Questions for authors

Please see weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.