Exploring the effectiveness of naive spatio-temporal exploits for depth completion

• Online Access: http://hdl.handle.net/2047/D20398280

With an increasing need for usable depth for autonomous navigation systems such as self-driving cars, depth completion is becoming an increasingly studied subject. RGB data provide much-needed aid in providing good recreation of dense depth maps from sparse LiDAR output. Yet, these data are also provided in sequential form. And thus,for this thesis, we aim to explore how effective using network layers that exploit Spatio-temporal features would be in achieving higher depth completion accuracy. We propose adding 3D convolutional layers and ConvGRU layers to a preexisting depth completion network and perform ablation studies on the effectiveness of these methods. We were able to verify that naive approaches can garner improvements quantitatively and qualitatively, but training results show that additional geometric constraints would perhaps boost such exploits even further for better depth completion results.