Multiscale feature fusion-based monocular depth estimation

Predicting a dense depth picture from a single color image is known as monocular depth estimation. A monocular depth estimate technique based on multi-scale feature fusion is suggested to address the shortcomings of the existing algorithms, which include inadequate feature fusion and a limited capacity to capture contextual information. A pyramid pooling-based global region improvement module enhances the network's feature extraction capabilities while providing rich global contextual information for scene depth analysis. The method extracts global features at several scales using a hierarchical Transformer encoder. To enhance the use of various scale feature information, the encoder's multi-scale feature output is chosen and fused by the feature selection fusion module, which also incorporates the encoder's underlying location feature information in the decoder. Experiments on the NYU Depth V2 dataset show that the method may significantly improve the precision of the monocular depth estimate, showing its effectiveness with a root mean square error (RMSE) of 0.341 and an accuracy (δ₁<1.25) of 0.915.