Abstract:
Transformer, which excels in capturing long-range dependencies, has shown great performance in a variety of computer vision tasks. In this paper, we propose a hybrid network with a Transformer-based encoder and a CNN-based decoder for monocular depth estimation. The encoder follows the architecture of classical Vision Transformer. To better exploit the potential of the Transformer encoder, we introduce the Attention Supervision to the Transformer layer, which enhances the representative ability. The down-sampling operations before the Transformer encoder lead to degradation of the details in the predicted depth map. Thus, we devise an Attention-based Up-sample Block and deploy it to compensate the texture features. Experiments on both indoor and outdoor datasets demonstrate that the proposed method achieves the state-of-the-art performance on both quantitative and qualitative evaluations. The source code and trained models can be downloaded at https://github.com/WJ-Chang-42/ASTransformer.
