Flow-CDNet: Dual-Branch Change Detection

• Flow-CDNet is a dual-branch neural network that integrates multi-scale optical flow detection with binary segmentation to capture both gradual and abrupt changes in bitemporal imagery.
• It employs a RAFT-like iterative module alongside ResNet50-based feature extraction to enhance geometric alignment and reduce false positives.
• The design uses a composite loss combining Tversky and L2 norms and introduces the FEPE metric to jointly assess segmentation accuracy and flow precision.

Flow-CDNet refers to a neural network architecture specifically designed to detect both slow and fast changes in bitemporal imagery, addressing the challenge of jointly modeling subtle gradual variations and abrupt discrete events. The approach integrates a multi-scale optical flow estimation module with a binary change detection stream, using mutual guidance and a composite loss to optimize for both continuous motion and sudden object-level alterations. Flow-CDNet introduces a specialized dataset, a composite loss based on binary Tversky and L2-norm, and a new evaluation metric called FEPE (F1-score over End-Point Error) for comprehensive assessment (2507.02307).

1. Problem Domain and Motivation

Change detection in bitemporal images involves localizing regions that have undergone alteration between two images taken of the same scene at different times. Many real-world applications—such as monitoring slopes, dams, and tailings ponds—require not only identification of significant (fast) changes but also the detection of weak, slow changes that could act as early indicators for hazards. Traditional architectures tend to treat change detection as either a motion estimation (optical flow) problem or as a binary segmentation task, but rarely both in concert. This duality motivates the Flow-CDNet design, which aims for robust performance across the spectrum of subtle to abrupt change (2507.02307).

2. Architectural Components

Flow-CDNet is organized in two mutually enhancing branches, each with a distinctive role and backbone:

2.1 Optical Flow Detection Branch (OFbranch)

• Based on a RAFT-like architecture, the OFbranch estimates dense, pixel-wise displacements between the earlier ($T_0^0$) and later ($T_0^1$) images.
• Features are extracted from both images via a shared feature encoder, followed by computation of a full 4D correlation volume $C$ over matching positions.
• A crucial multi-scale (pyramid) design applies average pooling with kernel sizes $\{1, 2, 4, 8\}$, yielding pyramids $\{C^1, C^2, C^3, C^4\}$. This enables the system to capture displacements ranging from large to extremely fine.
• Displacement refinement is performed iteratively using a convolutional GRU:

$f_{k+1} = f_k + \Delta f$

where $\Delta f$ is the update inferred from local neighborhood evidence in the multi-scale pyramid.

2.2 Binary Change Detection Branch (CDbranch)

• Focuses on discrete "fast" changes—objects emerging, disappearing, or fundamentally altered.
• The estimated flow from the OFbranch is used to warp $T_0^1$, aligning it to $T_0^0$’s viewpoint for better correspondence.
• The absolute difference between $T_0^0$ and the warped $T_0^1$ is computed, highlighting changing zones.
• Deep features are extracted from the difference image via a ResNet50 backbone. A spatial pyramid pooling module integrates context from various receptive fields.
• Upsampled features are concatenated and passed through further convolution/NL blocks, with a Sigmoid activation generating a binary change mask.

2.3 Branch Interaction

• The output of the OFbranch explicitly guides the CDbranch; optical flow-driven warping not only improves geometric alignment but also helps the CDbranch suppress false positives and detect subtle, motion-supported changes.
• Supervision provided by the CDbranch, in turn, modulates the focus of the OFbranch, particularly in ambiguous or low-motion regions.

3. Composite Loss Function

A dual-purpose loss function is crafted to train both branches simultaneously, balancing between motion estimation and binary change detection:

• L2 Norm Loss:

$$\text{loss}_{l2} = \lVert \text{output}_1 - \text{label}_1 \rVert_2 \cdot (1 - \text{label}_2)$$

Here, $\text{output}_1$ is the predicted flow, $\text{label}_1$ the ground-truth flow, and $(1 - \text{label}_2)$ masks out regions with fast changes.

• Binary Tversky Loss:

$$\text{loss}_{Tversky} = \frac{\text{masked}_{gt}}{\text{masked}_{gt} + \alpha \cdot \text{wrong}_{classified} + \beta \cdot \text{unmasked}_{gt}}$$

with: - $$\text{masked}_{gt} = \text{output}_2 \cdot \text{label}_2$$ - $$\text{wrong}_{classified} = (1 - \text{output}_2) \cdot \text{label}_2$$ - $$\text{unmasked}_{gt} = \text{output}_2 \cdot (1 - \text{label}_2)$$

• Total Loss:

$$\text{loss}_{total} = \text{loss}_{l2} + \psi \cdot \text{loss}_{Tversky}$$

with $\psi$ balancing the contribution of segmentation vs. flow accuracy.

This design ensures that slow-motion regions prioritize precise flow (through L2), while fast-changing regions weigh segmentation accuracy more heavily (through Tversky).

4. FEPE Evaluation Metric

The Flow-CDNet evaluation introduces FEPE (F1-score over End-Point Error), which integrates discrete change detection and motion estimation into a unified metric:

• F1-score: Calculated from predicted vs. reference binary segmentation of change regions.
• mEPE (mean End-Point Error): Average Euclidean distance between predicted and GT flows, measured over the union of all flagged regions.
• Formula:

$$\mathrm{FEPE} = \frac{F_1}{\mathrm{mEPE} + \epsilon}$$

where $\epsilon$ is a small constant to prevent division by zero.

FEPE jointly rewards networks that achieve high segmentation accuracy (F1) and low displacement error (mEPE), ensuring balanced optimization of both goals (2507.02307).

5. Empirical Results and Ablation Analysis

Extensive quantitative experiments were conducted on the purpose-built Flow-Change dataset:

• The RAFT-based Flow-CDNet variant achieved an F1-score of 0.892 for fast change detection, mEPE of 1.027 for flow accuracy, and a FEPE value of 0.869.
• Substitution of alternative flow backbones (LiteFlowNet, SpyNet) resulted in lower FEPE, underscoring the value of the multi-scale and iterative RAFT-like approach.
• Ablation studies demonstrate mutual improvement: disabling either branch degrades overall system performance in both motion and binary change domains, confirming synergetic effect.

Variant	F1-score	mEPE	FEPE
Flow-CDNet (RAFT)	0.892	1.027	0.869
Only OFbranch	—	>1.027	<0.869
Only CDbranch	<0.892	—	<0.869

(Values other than Flow-CDNet main result are summary indications.)

6. Application Domains and Dataset

The Flow-CDNet framework targets scenarios where both weak and overt changes must be characterized, such as:

• Geotechnical monitoring (slopes, dams, tailings ponds)
• Environmental change analysis
• Surveillance requiring both motion tracking and appearance-based differencing

The self-built Flow-Change dataset, constructed to support evaluation of both slow and fast changes, is integral for benchmarking such approaches, providing pixel-level ground truths for flow and change mask supervision (2507.02307).

7. Significance and Impact

Flow-CDNet establishes a dual-branch paradigm, demonstrating that joint exploitation of motion and appearance—each enhanced by multi-scale feature modeling and mutual supervision—significantly improves both the sensitivity to slow changes and the accuracy for explicit event detection. The composite loss and FEPE evaluation metric together set a comprehensive standard for future research in bitemporal scene understanding. Results show not only improved detection capability but also highlight the necessity of unified architectures for complex spatiotemporal change detection tasks in practical settings (2507.02307).