Multi-View Three-Dimensional Reconstruction based on a Two-Stage Multi-Level Depth Network for Agriculture Applications

To address the problems appearing in multi-view three-dimensional (3D) reconstruction, such as the improvement of the accuracy and completeness of the 3D reconstructed images, a two-stage multi-level depth network is proposed. In the stage 1 of the proposed network, several convolutional block attention modules (CBAMs) are applied in the lateral connections of the feature pyramid network (FPN). This is targeted to enhance the spatial and channel relativity of the different hierarchical feature maps so as to bring more semantic information. In the stage 2, the obtained multi-scale feature maps in the stage 1 are tackled by a set of cascaded processing procedures, such as adaptive propagation, single-trees transform, and matching cost computation. As a result, a depth map could be generated and then be further refined in the processing. Comparing with other state-of-the-art methods, the subjective and objective experiments based on the DTU dataset show that our method performs better result in completeness meanwhile maintaining a considerable overall metric. The investigation of applying the proposed method for reconstructing agricultural crop images was carried out, which is based on a set of self-collected images. The experiment shows that a suitable human visual perception for the images could be obtained.