BoostingDepth

This repository contains the source code of our paper: Guangkai Xu, Wei Yin, Hao Chen, Kai Cheng, Feng Zhao, Chunhua Shen, Towards 3D Scene Reconstruction from Locally Scale-Aligned Monocular Video Depth (Boosting Monocular Depth Estimation with Sparse Guided Points)

Prerequisite

conda create -n BoostingDepth python=3.7
conda activate BoostingDepth
pip install -r requirements.txt

Quick Start (Local recovery strategy)

1. (Optional) Run a demo inference.

   python lwlr.py
   

   RGB	GT depth	Pred depth global	Pred depth lwlr

   AbsRel: 0.079 –> 0.017

2. Prepare monocular depth prediction(e.g. LeReS) and sparse depth under test_imgs/. The sparse depth should have the same shape as the dense one, but fill with 0 where are invalid. Transfer them to .npy files, and organize as follows.

   |--test_imgs
   |   |--pred_depth_mono
   |   |   |--0.npy
   |   |   |--1.npy
   |   |   |--2.npy
   |   |--sparse_depth
   |   |   |--0.npy
   |   |   |--1.npy
   |   |   |--2.npy
   

3. Inference, the output can be seen under test_imgs/output_lwlr_depth/

   python inference_lwlr.py
   

Training & Inference (coming soon…)

GitHub

View Github