CNNapsule: A Lightweight Network with Fusion Features for Monocular Depth Estimation

Depth estimation from 2D images is a fundamental task for many applications, for example, robotics and 3D reconstruction. Because of the weak ability to perspective transformation, the existing CNN methods have limited generalization performance and large number of parameters. To solve these problems, we propose CNNapsule network for monocular depth estimation. Firstly, we extract CNN and Matrix Capsule features. Next, we propose a Fusion Block to combine the CNN with Matrix Capsule features. Then the skip connections are used to transmit the extracted and fused features. Moreover, we design the loss function with the consideration of long-tailed distribution, gradient and structural similarity. At last, we compare our method with the existing methods on NYU Depth V2 dataset. The experiment shows that our method has higher accuracy than the traditional methods and similar networks without pre-trained. Compared with the state-of-the-art, the trainable parameters of our method decrease by 65%. In the test experiment of images collected in the Internet and real images collected by mobile phone, the generalization performance of our method is further verified.