Self-distilled Feature Aggregation for Self-supervised Monocular Depth Estimation

Self-supervised monocular depth estimation has received much attention recently in computer vision. Most of the existing works literature aggregate multi-scale features for prediction via either straightforward concatenation or element-wise addition, however, such feature aggregation operations generally neglect contextual consistency between features. Addressing this problem, we propose Self-Distilled Feature Aggregation (SDFA) module simultaneously aggregating a pair low-scale and high-scale maintaining their consistency. The SDFA employs three branches to learn offset maps respectively: one map refining input other two under designed self-distillation manner. Then, an SDFA-based network self-supervised estimation, design self-distilled training strategy train proposed with module. Experimental results on KITTI dataset demonstrate that method outperforms comparative state-of-the-art methods most cases. code is available at https://github.com/ZM-Zhou/SDFA-Net_pytorch.