Zhong Wu

Existing stereo matching networks with high accuracy generally demand high computational overhead, making them expensive to deploy and unsuitable for many real-time applications, especially on source-constrained hardware platforms. Some recent works have tried to build lightweight stereo matching networks, but their accuracy and inference speed is far from satisfactory when deployed on source-constrained GPUs. This paper aims to build a lightweight stereo matching network with high accuracy. First, we claim that a reasonable disparity upsampling strategy is crucial for lightweight stereo matching to achieve high accuracy, whose importance is underestimated in existing works. Accordingly, an efficient and effective disparity upsampling strategy named Spatial Adaptive Disparity Shuffle (SADS) is proposed to meet the upsampling requirements of disparities within diverse regions. Second, a novel Channel-Disparity-Mixed Attention (CDMA) mechanism is proposed to regularize the high-resolution 4D compact cost volume. The attention values in CDMA are adaptive across channels, pixels and disparity candidates, making CDMA suitable for regularizing the compact cost volume. Based on the proposed methods, a lightweight stereo matching network (SADSNet) is designed that achieves real-time performance with high accuracy on source-limited hardware platforms (e.g., the NVIDIA Jetson TX2). In addition, SADSNet can be easily scaled up or down to suit different application scenarios according to available computational power. Extensive experimental results on multiple datasets show its clear superiority over SOTA lightweight stereo models and even many large accuracy-oriented models.

Few-shot classification algorithms have gradually emerged in recent years, and many breakthroughs have been made in the research of migration networks, metric spaces, and data enhancement. However, the few-shot classification algorithm based on Graph Neural Network is still being explored. In this paper, an edge-weight single-step memory-constraint network is proposed based on mining hidden features and optimizing the attention mechanism. According to the hidden distribution characteristics of edge-weight data, a new graph structure is designed, where node features are fused and updated to realize feature enrichment and full utilization of limited sample data. In addition, based on the convolution block attention mechanism, different integration methods of channel attention and spatial attention are proposed to help the model extract more meaningful features from samples through feature attention. The ablation experiments and comparative analysis of each training mode are carried out on standard datasets. The experimental results obtained prove the rationality and innovation of the proposed method.

Cost aggregation is crucial to the accuracy of stereo matching. A reasonable cost aggregation algorithm should aggregate costs within homogeneous regions where pixels have the same or similar disparities. Otherwise, the estimated disparity map is prone to the well‐known edge‐fattening issue and the problem of losing fine structures. However, current state‐of‐the‐art convolutional neural networks (CNNs) mainly do cost aggregation with square‐kernel convolutional layers that learn to adjust their kernel elements to make the actual receptive fields of the aggregated costs adapt to homogeneous regions with various shapes. This is a mechanism that easily leads to the above issues due to the translation equivalence and content‐insensitivity properties of CNNs. To tackle these problems, a novel densely connected asymmetric convolution block (Dense‐ACB) based on asymmetric convolutions is proposed to explicitly construct receptive fields with various shapes, which effectively alleviates the issues caused by mismatching shapes of receptive fields and homogeneous regions. The proposed Dense‐ACB brings new insight to CNN‐based stereo matching networks. Based on the proposed cost aggregation method, an efficient and effective stereo matching network is built, which not only achieves competitive overall accuracy compared with state‐of‐the‐art models but also effectively alleviates the edge‐fattening problem and preserves fine structures.