Cascaded Transposed Long-range Convolutions for Monocular Depth Estimation

Go Irie, Daiki Ikami, Takahito Kawanishi, Kunio Kashino; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020

We study the shape of the convolution kernels in the upsampling block for deep monocular depth estimation. First, our empirical analysis shows that the depth estimation accuracy can be improved consistently by only changing the shape of the two consecutive convolution layers with square kernels, e.g., (5 x 5) -> (5 x 5), to two "long-range" kernels, one having the transposed shape of the other, e.g., (1 x 25) -> (25 x 1). Second, based on this observation, we propose a new upsampling block called Cascaded Transposed Long-range Convolutions (CTLC) that uses parallel sequences of two long-range convolutions with different kernel shapes. Experiments with NYU Depth V2 and KITTI show that our CTLC offers higher accuracy with fewer parameters and FLOPs than state-of-the-art methods.