Multi-scale Feature Distribution Prediction Decoder for Lightweight Monocular Depth Estimation

Recently, various lightweight monocular depth estimation methods have been proposed. However, most methods mainly emphasize designing complex encoders while employing simple bilinear upsampling and convolution structures for decoders. That may degrade the performance of depth estimation. To address this problem, we propose a novel decoder based on image-guided multi-scale distribution prediction upsampler for lightweight monocular depth estimation. Specifically, through a distribution prediction block, our decoder predicts the distribution for every scale feature and upsamples them. The block fuses guidance and features to predict the distribution for features at higher resolutions. The feature at the lower resolutions are then upsampled to higher resolutions by a convolution-like structure, with the distribution serving as weights. Furthermore, our proposed decoder features an adaptive structure, allowing it to be combined with either single-scale or multi-scale encoders to form monocular depth estimation networks. In our experiments, we propose two networks by combining the decoder with single-scale and multi-scale encoders. Then supervised and self-supervised training of the two networks are conducted respectively. Experimental results show that our decoder has higher accuracy regardless of supervised or self-supervision training, and whether the encoder is single-scale or multi-scale. Moreover, compared with the simple decoders, our proposed decoder maintains a comparative number of parameters.