Abstract
Street Scene Change Detection (SSCD) aims to locate the changed regions between a given street-view image pair captured at different times, which is an important yet challenging task in the computer vision community. The intuitive way to solve the SSCD task is to fuse the extracted image feature pairs, and then directly measure the dissimilarity parts for producing a change map. Therefore, the key for the SSCD task is to design an effective feature fusion method that can improve the accuracy of the corresponding change maps. To this end, we present a novel Hierarchical Paired Channel Fusion Network (HPCFNet), which utilizes the adaptive fusion of paired feature channels. Specifically, the features of a given image pair are jointly extracted by a Siamese Convolutional Neural Network (SCNN) and hierarchically combined by exploring the fusion of channel pairs at multiple feature levels. In addition, based on the observation that the distribution of scene changes is diverse, we further propose a Multi-Part Feature Learning (MPFL) strategy to detect diverse changes. Based on the MPFL strategy, our framework achieves a novel approach to adapt to the scale and location diversities of the scene change regions. Extensive experiments on three public datasets (i.e., PCD, VL-CMU-CD and CDnet2014) demonstrate that the proposed framework achieves superior performance which outperforms other state-of-the-art methods with a considerable margin.
Introduction
Street Scene Change Detection (SSCD) involves identifying meaningful alterations between two images of the same location taken at different times—such as the appearance of a new vehicle or the removal of a road sign. This task is complicated by "noisy" changes like shifting shadows, seasonal variations, and camera viewpoint misalignments. Traditional methods often rely on naive feature fusion, which fails to bridge the semantic gap between different network layers. Furthermore, changes in street scenes are spatially unbalanced, varying significantly in scale and location.
To address these challenges, this paper introduces HPCFNet, a hierarchical Siamese network that adaptively fuses paired feature channels and employs a multi-part learning strategy to handle diverse scene changes.
Methodology
The HPCFNet architecture is built upon three primary innovations designed to improve the precision of change maps:
1. Paired Channel Fusion (PCF) Module
Rather than simple concatenation, the PCF module utilizes a Cross Feature Stack to interweave channels from the image pair. It then applies Parallel Atrous Group Convolution (PAGC). By using varying dilation rates within specific groups, the network captures multi-scale correlations between images efficiently, ensuring that semantic features are compared accurately across the hierarchy.
2. Reverse Spatial Attention (RSA)
To isolate changes from static backgrounds, the RSA module identifies commonalities between feature maps through multiplication. It then generates and reverses a similarity mask. This forces the model to suppress the static background and focus exclusively on regions where features do not match, making the changes "pop out."
3. Multi-Part Feature Learning (MPFL)
Street scene changes are often localized or occur at specific scales. The MPFL strategy partitions feature maps into four parallel branches—Global, Horizontal/Vertical, Local (quadrants), and Corner—to ensure the network is spatially aware and can detect changes regardless of their position in the frame.
Experiments and Results
The model was rigorously evaluated on three major benchmarks: PCD (Point Cloud Dataset), VL-CMU-CD (capturing extreme seasonal variations), and CDnet2014 (adapted for video-based change detection).
Quantitative Performance
HPCFNet established new state-of-the-art (SOTA) results across all datasets:
- PCD Dataset: Achieved an F-Score of 0.776 on the GSV subset and 0.868 on TSUNAMI, marking a significant improvement over previous methods like CSCDNet.
- VL-CMU-CD: Reached an F-Score of 0.752, outperforming the nearest competitor by 4.2%.
Qualitative Analysis
Visual results demonstrate that HPCFNet is exceptionally capable of resolving fine details. While previous models often blurred small objects or missed slender structures like poles, HPCFNet accurately delineates these changes even under challenging lighting conditions.
Conclusion
HPCFNet advances street scene change detection by moving beyond naive fusion toward a hierarchical, channel-aware comparison. The combination of the Reverse Spatial Attention mechanism and the Multi-Part Feature Learning strategy allows the model to overcome traditional biases toward the center of the image and ignore environmental noise. Future work may explore lighter backbones for real-time applications and the integration of temporal consistency for video sequences.
Summary updated