We have implemented a neural-network-based depth estimation solution with two different models. User can also choose different colour maps to view the estimated depth.
Depth estimation is implemented using a model trained on two different dataset: KITTI and NYUDepth v2 dataset. The depth estimation model consists of adensenet169encoder and a decoder withone 2D-convolution layer, then four upsampling layers, followed by another 2D-convolution layer. Weused pre-trained models that were operated under keras and converted them into PyTorchptmodels.The output is a depth map with a dimension ofw2byh2, wherewandhare the original size of theinput image. In order to display the depth map, we used different color schemes provided in the librarymahdilamb:colormap.
See the project report for more details.