CNN-MonoFusion

The release code and dataset of CNN-MonoFusion for ismar2018.

Introduction

The project contain two submodules (depth-esti/pointcloud-fusion).

• Network-Archi can be found at adenet_def.py. We name our network as adenet (adaptive-depth-estimation-network), which combined the resnet50/astrous/concat layers and trained using our adaptive-berhu loss.

• For online depth prediction, you need run adenet_run_as_server.py in your server.

• The pointcloud-fusion is used for fusion stage. We build our system depended on a mono-slam system, so you need to incorpolate our fusion code into a slam sysytem like ORB-SLAM.

Requirements

We run our whole system in Win7&Win10 locally ok.
You can also running the network on the server compute for accerlating the speed!

• Tensorflow >= 1.4.0 & CUDA>=8.0
• Python 3.5.2
• CMake 3.6.0
• Visual Studio 2015
• Optional: kinect-v2 for collecting your own dataset

Dataset

All the images are collected by NetEaseAI-CVLab.
Copyright @2018 CNN-MonoFusion Authors. All rights reserved.
Please download the NEAIR-dataset here.

Pretrained-Models

You can download our models used in the paper here

Citing CNN_MonoFusion (update later)

If you find CNN_MonoFusion useful in your research, please consider citing:

@article{
    Author = {Jiafang Wang, Haiwei Liu, Lin Cong, Zuoxin Xiahou, and Liming Wang},
    Title = {CNN-MonoFusion: Online Monocular Dense Reconstruction using Learned Depth from Single View},
    Journal = {},
    Year = {}
}