GLTB Depth: Monocular Depth Estimation Based on Global-Local Transformer Black

The encoder-decoder architecture-based models are widely used for estimate of monocular depth. The topologies of deep neural networks are constantly evolving, the model architecture has become increasingly complex, and the number of model parameters has significantly increased. This results in excessive memory usage, making inference on low-end devices challenging. Additionally, traditional lightweight encoders perform poorly, and the conventional encoder-decoder-based models suffer from insufficient feature extraction and an inability to adapt to spatial deformations of objects. To address these issues, this study suggests a brand-new monocular depth estimation technique based on the ResNet-50 lightweight encoder. First, a Gaussian Dynamic Dense Atrous Spatial Pyramid Pooling (GDASPP) feature fusion method is introduced. Then, to make up for the encoder's inadequate feature extraction, a Global-Local Transformer Block (GLTB) mechanism is proposed. Finally, the decoder uses Deformable Convolutional Networks (DCN) to adjust to object spatial deformations. Results from experiments on the KITTI outdoor scene dataset demonstrate that our approach significantly improves performance.