Abstract
Optical flow estimation in low-light, high-noise environments—termed Dark Optical Flow Estimation (DOFE)—is a significant challenge due to the breakdown of brightness constancy and the presence of sensor noise. We present CEDFlow, a framework that enhances latent features rather than raw pixels. By decomposing features into spatial frequencies and applying a 2nd-order Gaussian difference ($D^2oG$) operator, CEDFlow successfully isolates salient contours from noise. Our method achieves state-of-the-art results on major dark-scene benchmarks, including a 1.08 EPE on the FCDN dataset.
Introduction
Optical flow is fundamental to computer vision tasks like autonomous driving and video editing. Traditional methods rely on the assumption that a pixel's intensity remains constant across frames. However, in dark environments, low contrast and high sensor noise violate this assumption, leading to "fake" textures and blurred boundaries. While image enhancement can help, it often introduces artifacts detrimental to flow tracking. CEDFlow addresses this by performing enhancement directly in the feature (latent) space, prioritizing structural integrity over visual aesthetics.
Methodology
The CEDFlow architecture introduces a multi-stage approach to robust feature extraction in the dark:
- Spatial Frequency Decomposition: Input features are split into high-frequency (fine details/noise) and low-frequency (broad shapes) components.
- Dual-branch Motion Encoder (DBME): A Global Encoder processes low-frequency data to capture long-range motion, while a Local Encoder handles high-frequency structural details. These branches are fused via a Multi-layer Perceptron (MLP).
- Latent Contour Enhancement (LCE): The core innovation is the application of a 2nd-order Gaussian difference ($D^2oG$) operator in the latent space. This acts as a band-pass filter that suppresses high-frequency noise while amplifying large-scale, salient contours.
- Refinement: A sigmoid-based double-pass filter further suppresses residual noise before the features enter the final flow update module, which is based on the RAFT architecture.
Experiments and Results
CEDFlow was evaluated on synthetic (FCDN) and real-world (VBOF, FLIR ADAS) datasets.
Quantitative Performance
On the FCDN (FlyingChairs Dark&Noise) benchmark, CEDFlow achieved an End-Point Error (EPE) of 1.08, outperforming AGFlow (1.15) and the baseline RAFT (1.23). On the real-world VBOF "Fuji2" subset, it reduced error to 13.94, a 3.2% improvement over the next best method.
Qualitative Analysis
Visual results demonstrate that CEDFlow maintains sharp object boundaries where other methods produce jagged or blurry flow maps. By focusing on the "skeleton" of the scene through the $D^2oG$ operator, the model remains stable even under extreme noise.
Conclusion
CEDFlow demonstrates that latent-space enhancement is a powerful strategy for dark-scene vision tasks. By treating local and global motion separately and focusing on large-scale contours, it overcomes the limitations of pixel-level enhancement. Future work will focus on optimizing the computational overhead (currently 7.7M parameters) and exploring applications in low-light object detection and depth estimation.
Summary updated