Deep learning has recently been proven to deliver excellent performance in multi-view stereo (MVS). However, it is difficult for deep learning-based MVS approaches to balance their efficiency and effectiveness. Towards this end, we propose the DSC-MVSNet, a novel coarse-to-fine and end-to-end framework for more efficient and more accurate depth estimation in MVS. In particular, we propose an attention aware 3D UNet-shape network, which first uses the depthwise separable convolutions for cost volume regularization. This mechanism enables effective aggregation of information and significantly reduces the model parameters and computation by transforming the ordinary convolution on cost volume as depthwise convolution and pointwise convolution. Besides, a 3D-Attention module is proposed to alleviate the feature mismatching problem in cost volume regularization and aggregate the important information of cost volume in three dimensions (i.e. channel, space, and depth). Moreover, we propose an efficient Feature Transfer Module to upsample the low-resolution (LR) depth map to a high-resolution (HR) depth map to achieve higher accuracy. With extensive experiments on two benchmark datasets, i.e. DTU and Tanks & Temples, we demonstrate that the parameters of our model are significantly reduced to $$25\%$$25 % of the state-of-the-art model MVSNet. Besides, our method outperforms or maintains on par accuracy with the state-of-the-art models. Our source code is available at https://github.com/zs670980918/DSC-MVSNet.

Satellite multi-view stereo (MVS) is a fundamental task in large-scale Earth surface reconstruction. Recently, learning-based multi-view stereo methods have shown promising results in this field. However, these methods are mainly developed by transferring the general learning-based MVS framework to satellite imagery, which lacks consideration of the specific terrain features of the Earth’s surface and results in inadequate accuracy. In addition, mainstream learning-based methods mainly use equal height interval partition, which insufficiently utilizes the height hypothesis surface, resulting in inaccurate height estimation. To address these challenges, we propose an end-to-end terrain feature-aware height estimation network named SA-SatMVS for large-scale Earth surface multi-view stereo, which integrates information across different scales. Firstly, we transform the Sobel operator into slope feature-aware kernels to extract terrain features, and a dual encoder–decoder architecture with residual blocks is applied to incorporate slope information and geometric structural characteristics to guide the reconstruction process. Secondly, we introduce a pixel-wise unequal interval partition method using a Laplacian distribution based on the probability volume obtained from other scales, resulting in more accurate height hypotheses for height estimation. Thirdly, we apply an adaptive spatial feature extraction network to search for the optimal fusion method for feature maps at different scales. Extensive experiments on the WHU-TLC dataset also demonstrate that our proposed model achieves the best MAE metric of 1.875 and an RMSE metric of 3.785, which constitutes a state-of-the-art performance.