Visualization of Convolutional Neural Networks for Monocular Depth Estimation

Junjie Hu, Yan Zhang, Takayuki Okatani http://arxiv.org/abs/1904.03380

Introduction

We attempt to interpret CNNs for monocular depth estimation. To this end, we propose to locate the most relevant pixels of input image to depth inference. We formulate it as an optimization problem of identifying the smallest number of image pixels from which the CNN can estimate a depth map with the minimum difference from the estimate from the entire image.

Predicted Masks

Extensive experimental results show

• The behaviour of CNNs that they seem to select edges in input images depending not on their strengths but on importance for inference of scene geometry.

• The tendency of attending not only on the boundary but the inside region of each individual object.

• The importance of image regions around the vanishing points for depth estimation on outdoor scenes.

Please check our paper for more details.

Dependencies

• python 2.7
  
• pytorch 3.1
  

Running

Download the trained networks for depth estimation : https://drive.google.com/file/d/1QaUkdOiGpMuzMeWCGbey0sT0wXY0xtsj/view?usp=sharing

Download the trained networks for mask prediction : https://drive.google.com/file/d/12VXcfSEZ_Te13w4WYJyoaJjkhUxOEqwo/view?usp=sharing

Download the NYU-v2 dataset: https://drive.google.com/file/d/1WoOZOBpOWfmwe7bknWS5PMUCLBPFKTOw/view?usp=sharing

• Test
  

  python test.py
  

• Training
  

  python train.py
  

Citation

If you use the code or the pre-processed data, please cite:

@article{hu2019visualization,
  title={Visualization of Convolutional Neural Networks for Monocular Depth Estimation},
  author={Hu, Junjie and Zhang, Yan and Okatani, Takayuki},
  journal={arXiv preprint arXiv:1904.03380},
  year={2019}
}