Depth Estimation from Single Image Using CNN-Residual Network

In this project, we tackle the problem of depth estimation from single image. The mapping between a single image and the depth map is inherently ambiguous, and requires both global and local information. We employ a fully convolutional architecture, which first extracts image feature by pretrained ResNet-50 network. We do transfer learning by replacing the fully connected layer of ResNet-50 with upsampling blocks to recover the size of depth map. The upsampling block combines residual learning concept. This CNN-Residual network can be trained en-to-end, and runs real time on images with enough computing power. We demonstrate that our method of doing up-sampling by CNN-Residual network yields better result than fully connected layer, because it avoids overfitting. We also compare our model with pure CNN network and illustrates the effectiveness of tansfer learning. We also show the influence of different loss functions during training. The results are shown by comparing qualitative visualization and quantitative metrics.