Skip to content

Latest commit

 

History

History
48 lines (36 loc) · 2.42 KB

README.md

File metadata and controls

48 lines (36 loc) · 2.42 KB
Raw

Conf-Net: Depth Completion with Error-Map

Tensorflow implementation of our paper Conf-Net: Predicting Depth Completion Error-Map For High-Confidence Dense 3D Point-Cloud.

Introduction

This work proposes a new method for depth completion of sparse LiDAR data using a convolutional neural network which learns to generate ”almost” full 3D point-clouds with significantly lower root mean squared error (RMSE) over state-of-the-art methods. Our main contributions are listed below:

  • We propose a novel method to predict a high-quality pixel-wise error-map. Our approach outperforms existing methods in terms of uncertainty and confidence maps.
  • Our approach generates industry-level clean (high confidence - low variance) 360 ◦ 3D dense point-cloud from sparse LiDAR point-cloud. Our point-cloud is 15 times denser than input (which is Velodyne HDL 64 point-cloud) and 3 times more accurate than the state-of-the-art (RMSE = 300mm).
  • We conduct the uncertainty based analysis of Kitti depth completion dataset for the first time.

See the full demo on Youtube.

Network input/outputs

Purged 3D point-cloud

Sparse Depth (Input):

Predicted Dnese Depth (RMSE: 1000mm):

Predicted Pixelwise Error-Map:

Purged Dense Depth (RMSE: 300mm):

Results on Monocular Depth Estimation on NYU Depth V2:

Installing

The code is tested with Cuda 9 and tensorflow 1.10 on Titan V.
We suggest using our docker to run the code.
(The docker includes Cuda 9 with tensorflow 1.10, ros kinetict and pcl library for point-cloud visualization.)

bash docker_run.sh

Training

python ./main/depth_completion.py train

Testing

python ./main/deploy_haval.py