Deeper Depth Prediction with Fully Convolutional Residual Networks[3DV 16]

One Line Summary

Depth estimation using a fully convolutional architecture optimized using a reverse Huber loss without any post processing steps, showing the state of the art result on benchmark datasets.

Motivation

Encoder-decoder architectures uses the high level and low level features for optimal predictions, since depth estimation requires global and local uses, this architecture is well for suited depth estimation task.

Detailed Summary

• A novel approach to the problem of depth estimation from a single image, single-scale CNN architecture that follows residual learning. The proposed network is fully convolutional, comprising up-projection layers that allow for training much deeper configurations, illustrates a faster and more efficient approach to up-convolutional layers, berHu loss is used for the better convergence.

Novelty and Contributions

• Reverse Huber loss is introduced(berHu loss)
• Faster up convolutions, up projections are a faster implementaiton of up convolutions

Network Details

• Network uses the encoder-decoder architecture to capture both high level features and the low level features.
• ResNet 50 is used as the encoder part of the network .
• Decoder part contains the up projections.

Results

Authours

Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, Nassir Navab

Sources

Paper

Code