Stereo matching has been widely used in various computer vision applications and it is still a challenging problem. Adaptive support weights (ASW) methods represent the state-of-the-art in stereo matching and have achieved outstanding performance. However, the local ASW methods fail to resolve the matching ambiguity in low texture areas because their cost aggregation is limited within local fixed or adaptive support windows. On the other hand, the non-local ASW methods perform cost aggregation along a special tree, so that these methods are often sensitive to high texture areas since some useful connectivity constrains between adjacent pixels are broken during constructing the special tree. To solve these problems, in this paper, a novel and generic fusing ASW framework is proposed for stereo matching. In this framework, we establish dual support windows for each pixel, i.e., a local window and the whole image window. As such, the primitive connectivity between each pixel and its neighboring pixels in the local window can be maintained, and then each pixel not only gets appropriate supports from neighboring pixels within its local support window, but also receives more adaptive supports from the other pixels outside the local window. Furthermore, a local edge-aware filter and a non-local edge-aware filter, whose kernel windows correspond to the dual support windows, are merged in order to achieve collaborative filtering of the cost volume. Performance evaluation on the Middlebury and KITTI datasets shows that the proposed stereo matching method outperforms the current state-of-the-art methods.

Visually sensitive regions in the scene are thought to be important for scene categorization. In this paper, we propose to utilize the important visually sensitive information represented by deep features for scene categorization. Specifically, the context relationship between the objects and the surroundings is fully utilized as the main basis for judging the content of the scene, and combining with the deep convolution neural networks (CNNs), a scene categorization model based on deep visually sensitive features is constructed. First, the saliency regions of the scene images are marked according to the context-based saliency detection algorithm. Then, the original images and the corresponding visually sensitive region detection images are superimposed to obtain the visually sensitive region enhancement images. Second, the deep convolution features of the original images, the visually sensitive region detection images, and the visually sensitive region enhancement images are extracted through the deep CNNs pre-trained on the large-scale scene dataset Places. Finally, considering that the deep features extracted by different layers of the convolution network have different capabilities of discrimination, the fusion features are generated from multiple convolution layers to construct visually sensitive CNN model (VS-CNN). In order to verify the effectiveness of the proposed model, the experiments are conducted on the five standard scene datasets, i.e., LabelMe, UIUC-Sports, Scene-15, MIT67, and SUN. The experimental results show that the proposed model is effective and has good adaptability. Especially, our categorization performance is superior to many state-of the-art methods for a complex indoor scene.

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.