Current version: 1.0
ENFT (Efficient Non-consecutive Feature Tracking) is a feature tracking method which can efficiently match feature point correspondences among one or multiple video sequences. Our framework consists of steps of solving the feature ‘dropout’ problem when indistinctive structures, noise or even large image distortion exist, and of rapidly recognizing and joining common features located in different subsequences or even in different videos. This source code provides an efficient GPU implementation.
[1] Guofeng Zhang*, Haomin Liu, Zilong Dong, Jiaya Jia, Tien-Tsin Wong, and Hujun Bao*. Efficient Non-Consecutive Feature Tracking for Robust Structure-from-Motion. IEEE Transactions on Image Processing, 25(12): 5957 – 5970, 2016. [arXiv report][video]
[2] Guofeng Zhang, Zilong Dong, Jiaya Jia, Tien-Tsin Wong, and Hujun Bao. Efficient Non-Consecutive Feature Tracking for Structure-from-Motion. European Conference on Computer Vision (ECCV), 2010.
This source code only provides the feature tracking module. The whole executable SfM system and more datasets can be found at http://www.zjucvg.net/ls-acts/ls-acts.html.
This software is for non-commercial use only. Any modification based on this work must be open source and prohibited for commercial use.
If you need a closed-source version of ENFT for commercial purposes, please contact Guofeng Zhang.
If you use this source code for your academic publication, please cite our TIP paper:
@article{
title={Efficient Non-Consecutive Feature Tracking for Robust Structure-from-Motion},
author={Guofeng Zhang, Haomin Liu, Zilong Dong, Jiaya Jia, Tien-Tsin Wong, Hujun Bao},
journal={IEEE Transactions on Image Processing},
volume = {25},
number = {12},
papges = {5957--5970},
doi = {10.1109/TIP.2016.2607425},
year={2016}
}
The project is built by VS2010. All the dependent libraries must be built and linked. We also provide the x64 libraries built by VS2010 in lib/.
The entry to the program is in ENFT/ENFT.cpp. Here is an example:
int _tmain(int argc, _TCHAR* argv[])
{
std::vector<Video> videos(2);
videos[0].imgFileName = "../data/0/0000.jpg";
videos[0].start = 10;
videos[0].step = 2;
videos[0].end = 150;
videos[1].imgFileName = "../data/1/0000.jpg";
videos[1].start = 0;
videos[1].step = 2;
videos[1].end = 100;
RunENFT(videos, "./param/", "../data/result.txt");
return 0;
}
Video is a struct specifying the input video sequence.
struct Video
{
char *imgFileName;
int start, step, stop;
} Video;
In this example, inputs are two video sequences. The first video sequence consists of ../data/0/0010.jpg, ../data/0/0012.jpg, ..., ../data/0/0150.jpg and the second consists of ../data/1/0000.jpg, ../data/1/0002.jpg, ..., ../data/1/0100.jpg. We call RunENFT to run the algorithm:
void RunENFT(const std::vector<Video> &videos, const char *paramDir, const char *outputFileName);
The first argument is a vector of input video sequences. The second argument paramDir specifies a directory containing the files of all the parameters the algorithm needs. The third argument outputFileName specifies the output file saving the feature tracking result. The output file containts multiple lines, one feature track for each line. Each line starts with an integer N, followed by N feature correspondences. Each correspondence is defined by iSeq iFrm x y. The two integers iSeq and iFrm are respectively the index of video sequence and image frame in the sequence, and the two floats (x, y) is the feature location in the image.