Monitored Distillation for Positive Congruent Depth Completion

Unsupervised Depth Completion Guided by Visual Inertial System and Confidence

“…Our method is superior to the most advanced unsupervised depth completion methods and some supervised depth completion methods, and slightly worse than the most advanced supervised depth completion methods. In addition, compared with the most advanced unsupervised MDPC method [ 32 ] in Table 2 , our method can improve the MAE by 1.93%, RMSE by 1.76%, iMAE by 3.26% and iRMSE by 3.79%. It is worth emphasizing that the sparse depth input to our network is generated by VI-SLAM, which only accounts for 0.5% of the pixel density of the images, while other methods use the sparse depth maps with the built-in density of 5% in the dataset.…”

“…We submitted the proposed method to the KITTI Depth benchmark and compared it with advanced unsupervised and supervised methods. Among them, the MDPC method proposed by Liu et al is an adaptive knowledge distillation method, which integrates some existing excellent models while avoiding some errors in the models [ 32 ]. Among the supervised depth completion methods, the methods with higher accuracy are the MFF-Net method by Liu et al [ 33 ] and the NLSPN method by Park et al [ 12 ].…”

“…To display the improvements from our contributions, we show a positive qualitative comparison with advanced unsupervised method [ 9 , 32 ] in Figure 6 . The first line is RGB images, the second line is sparse depth maps generated by VI-SLAM (the size of depth is represented by different colors), the third line is KBNet method proposed by wong et al, the fourth line shows the MDPC method proposed by Liu et al, the fifth line shows the local magnification comparison map, and the sixth line shows our method, where challenging regions are highlighted.…”

“…In order to further verify the generalization ability of our method in real scenes, we, respectively, applied the model of our method and the literature [ 9 , 32 ] trained in VOID dataset to the depth completion of OpenLORIS dataset and compared them. We selected 1000 images from three complex real scenes of cafe, home, and market as the test, and used the provided IMU data and RGB images to generate sparse depth images with a density of 0.5%, ensuring the same input sources for the three methods.…”

SPNet: Structure preserving network for depth completion

“…Multi-modal fusion strategy can be roughly divided into three ways: early, middle and late fusion. Specifically, the early fusion [4,18,19,21,51] simply concatenated two modalities and then directly fed them into the same encoder. This fusion method did not fully consider the distinction of different modalities.…”