This repository contains the code for "FastFusion: Deep Stereo-LiDAR Fusion for Real-time High-precision Dense Depth Sensing"
Input image
Coarse disparity map
Refinement disparity map
Input image
Coarse disparity map
Refinement disparity map
This code is tested on:
- Ubuntu 20.04
- python 3.8
- pytorch 1.8.1
- torchvision 0.9.1
- opencv-python 4.5
- cuda 10.2
Download the KITTI 141 dataset which contains the selected 141 image pairs with corresponding different lines of LiDAR maps.
Download the coarse estimation from StereoBit with different lines of LiDAR input.
Move two directories into the "dataset" directory which should be like this:
dataset
│
└───pred
└───kitti141
Hyperparameters can be set in utils/option.py. To train a refinement network, one should first try to pre-train the refinement network in the scene flow dataset:
python train_SF.py
Fine-tuning the model on the KITTI 2015 dataset is also easy, you need to specify the path of the pre-trained model and then run:
python train_kitti2017.py
To evaluate the accuracy of the refinement network on misaligned raw Lidar signal, one can run:
python eva_kitti141.py