Skip to content

Latest commit

 

History

History
83 lines (70 loc) · 4.55 KB

README.md

File metadata and controls

83 lines (70 loc) · 4.55 KB
Raw

Multi-Scale Continuous CRFs as Sequential Deep Networks for Monocular Depth Estimation

By Dan Xu, Elisa Ricci, Wanli Ouyang, Xiaogang Wang and Nicu Sebe

Introduction

CCRF-CNN is a continuous CRFs model implemented with neural networks for structured fusion of multi-scale predictions which is applied in monocular depth estimation and was accepted at CVPR 2017.
The currently published version implements a multi-scale cascade continuous CRF model. The model is implemented as a CNN layer and can be also applicable to other continuous regression problems.
The code is implemented under Caffe and has been tested under the configurations of Ubuntu 14.04 and CUDA 8.0.
Links: [CVPR Paper][TPAMI Paper][Oral Presentation]

Trained Model, Deploy file and Testing Script

To take a practice of our method, please download a trained model (ResNet50 as front-end) on the standard training set and the deploy network file from Google Drive.

The structured fusion of multiple Scales of the predictions are performed using the developed MultiStageMeanField Caffe layer together with the continuous mean-field updating implementation. The number of mean-field interations could be specified, and the caffe prototxt definition is as follows:

##Structured fusion of two scales using Continous-CRF-CNN##
layer {
  name: "inference1" # set the name to be the same as in the training phase to 
  load parameters in the testing phase #
  type: "MultiContinuousMeanfield" bottom: "upscore-map3" bottom: "upscore-map4" bottom: "data" 
  top: "predicted-map1"
  param { lr_mult: 10000 # learning rate for weight of smoothness }
  param { lr_mult: 10000 # learning rate for weight of appearance }
  multi_stage_meanfield_param { num_iterations: 6 theta_alpha: 30 theta_beta: 15 theta_gamma: 3 
  spatial_weight: 3 bilateral_weight: 5 } 
  }
  
##Structured fusion of two scales using Continous-CRF-CNN (the output of inference 1 
will be used as an input scale in this phase)##
layer {
  name: "inference2" #if you set name "inference1", code will load parameters from caffemodel#
  type: "MultiContinuousMeanfield" bottom: "predicted-map1" bottom: "upscore-map5" bottom: "data" 
  top: "predicted-map2"
  param { lr_mult: 10000 # learning rate for weight of smoothness }
  param { lr_mult: 10000 # learning rate for weight of appearance }
  multi_stage_meanfield_param { num_iterations: 6 theta_alpha: 30 theta_beta: 15 theta_gamma: 3 
  spatial_weight: 3 bilateral_weight: 5 } 
  }
  
 ... To stack more for fusing more scales ...

An example testing script can be found from examples/CCRF_CNN_DepthEstimation/nyud_test.py.

Results on NYUD-V2

The results with ResNet-50 as a backbone and trained on the standard training set with 795 training images. Examples as follows:

Full test results can be directly downloaded from Here.

Results on Make3D

Results on Make3D test dataset can be downloaded from Here. The groundtruth depth map can be downloaded from Here.

Results on KITTI

Results on KITTI test dataset (on the Eigen split) can be downloaded from Here.

Citation

Please consider citing the following papers if the code is helpful in your research work:

@inproceedings{xu2017multi,
  title={Multi-Scale Continuous CRFs as Sequential Deep Networks for Monocular Depth Estimation},
  author={Xu, Dan and Ricci, Elisa and Ouyang, Wanli and Wang, Xiaogang and Sebe, Nicu},
  booktitle={CVPR},
  year={2017}
}
@article{xu2018monocular,
  title={Monocular Depth Estimation using Multi-Scale Continuous CRFs as Sequential Deep Networks},
  author={Xu, Dan and Ricci, Elisa and Ouyang, Wanli and Wang, Xiaogang and Sebe, Nicu},
  journal={IEEE TPAMI},
  year={2018}
}