Skip to content

Latest commit

 

History

History
290 lines (252 loc) · 7.27 KB

MANUAL.md

File metadata and controls

290 lines (252 loc) · 7.27 KB
Raw

Contents

Usage

SfmDataGenerator <command> -<param1>=<value1> -<param2>=<value2> ... <param_n>=<value_n>

Where <command> is one of:

In the following commands description SfmDataGenerator is omitted

Commands

help

Aliases: h, ?

Show help

Usage:

help

generate

Aliases: gen, g

Run dataset generation

Usage:

gen -in=<model> -tex=<texture> -out=<out-dir> [-format=<fmt>]  [-mode=<mode>]
Parameter Description Values Default
<model> Path to input model
<out-dir> Output directory
<fmt> Format of dataset txt, xml, yml txt
<mode> Dataset generation mode (SfM, Depth, Silhouette) sfm, dep, sil sfm

Note: input model is shifted and scaled to be in cube (-0.5, -0.5, -0.5) - (0.5, 0.5, 0.5)

view

Alias: v

View generated dataset

Usage:

view -in=<sfm-data> [-i1=<first-idx>] [-i2=<second-idx>]
Parameter Description
<sfm-data> Path to sfmData file
<first-idx> First image index
<second-idx> Second image index

Effect of view depends on parameters passed as well as type of dataset.

Dataset type i1 i2 Effect
SfM - - View point cloud & camera poses
SfM + - View features of (i1)-th image
SfM + + View matches between (i1)-th & (i2)-th images
RGB-D & Sil - - View camera poses
RGB-D & Sil + - View (i1)-th depth (or silhouette) image

convert

Aliases: conv, c

Convert dataset from one format to another. It uses file extension to determine format.

Usage:

conv -in=<input> -out=<output>
Parameter Description
<input> Input sfmData file
<output> Output sfmData file

noise-proj

Note: only for SfM datasets

Aliases: np, n

Add gaussian noise to each projection coordinates

Usage:

np -in=<input> -out=<output> [-stdev=<stdev>]
Parameter Description Default
<input> Input sfmData file
<output> File to save noisy dataset
<stdev> Strandard deviation of gaussian noise (in pixels) 0.03

false-proj

Note: only for SfM datasets

Aliases: fp, f

Add outliers to point-projection correspondences. It simulates wrong feature match.

Usage:

fp -in=<input> -out=<output> [-ratio=<ratio>] [-count=<count>]
Parameter Description Default
<input> Input sfmData file
<output> File to save noisy dataset
<ratio> Ratio of wrong correspondences 0.2
<count> Number of wrong correspondences 50

If <ratio> and <count> are both given, <ratio> has priority.

run

Alias: r

Run SfmDataGenerator in interactive mode

Interactive mode

Interactive mode was implemented to operate SfmDataGenerator via GUI.

Controls

Output formats

Output folder structure

For SfM and Silhouette datasets

output_dir
|
|--images
|  |--000001.png
|  |--000002.png
|  |-- ...
|--sfmData{.txt, .xml, .yml}

For RGB-D Datasets

output_dir
|
|--images
|  |--000001.exr (depth floating point image in OpenEXR)
|  |--000001.png (corresponding RGB image)
|  |--000002.exr
|  |--000002.png
|  |-- ...
|--sfmData{.txt, .xml, .yml}

sfmData structure

.txt

<nImages> <nPoints>
<path-to-image_1>                           (In case of RGB-D dataset this path references to a depth image)
<K_11> <K_12> <K_13> <K_21> ... <K_33>      (3x3 camera intrinsics matrix row by row)
<R_11> <R_12> <R_13> <R_21> ... <R_33>      (3x3 camera rotation matrix row by row)
<T_x> <T_y> <T_z>                           (camera translation vector)
<k_1> <k_2>                                 (Coefficients of radial distortion: k_1 for the second-order distortion and k_2 for the fourth-order distortion)
<nPointsOnImage_1>                            (Number of features)
<3d-point-index_1> <x_1> <y_1>              (Index of 3d point (from 0 to nPoints-1) and its projection)
<3d-point-index_2> <x_2> <y_2>
...
<3d-point-index_n> <x_n> <y_n>
<path-to-image_2>
...
<X_0> <Y_0> <Z_0>                           (0-indexed point cloud)
<X_1> <Y_1> <Z_1>
...

Note: for silhouette & RGB-D datasets <nPoints>, <nPointsOnImage> are zeros, i.e. only camera intrinsics & extrinsics are presented

.yml

%YAML:1.0
views:
    -
        image: <path-to-image_1>
        camera:
            - <K_11>
            - <K_12>
            ...
            - <K_33>
            - <R_11>
            - <R_12>
            ...
            - <R_33>
            - <T_x>
            - <T_y>
            - <T_z>
            - <k_1>
            - <k_2>
        observations:
            - <3d-point-index_1>
            - <x_1>
            - <y_1>
            - <3d-point-index_2>
            - <x_2>
            - <y_2>
            ...
    -
        image: <path-to-image_2>
        ...
    ...
points:
   - <X_0>
   - <Y_0>
   - <Z_0>
   - <X_1>
   - <Y_1>
   - <Z_1>
   ...

.xml

<?xml version="1.0"?>
<opencv_storage>
<views>
  <_>
    <image>path-to-image_1</image>
    <camera>K_11 K_12 ... R_11 ... R_33 ... k_1 k_2</camera>
    <observations>
      3d-point-index_1 x_1 y_1
      3d-point-index_2 x_2 y_2
      ...
    </observations>
  </_>
  ...
</views>
<points>
  X_0 Y_0 Z_0 X_1 Y_1 Z_1 ...
</points>
</opencv_storage>

Reading .xml or .yml datasets with OpenCV (from SfmDataGenerator source)

bool SfmData::loadFromXmlOrYml(const string &filename) {
	FileStorage fs(filename, FileStorage::READ);
	if (!fs.isOpened())
		return false;
	clear();
	for (auto &node : fs["views"]) {
		View view;
		Camera cam;
		string imgFile;
		node["image"] >> imgFile;
		node["camera"] >> cam;
		int it = 0;
		auto obsNode = node["observations"];
		for (auto it = obsNode.begin(); it != obsNode.end(); ++it) {
			int d;
			double x, y;
			(*it) >> d;
			++it;
			(*it) >> x;
			++it;
			(*it) >> y;
			view.emplace_back(d, x, y);
		}
		addView(cam, view, imgFile);
	}
	auto pointsNode = fs["points"];
	for (auto it = pointsNode.begin(); it != pointsNode.end(); ++it) {
		double x, y, z;
		(*it) >> x;
		++it;
		(*it) >> y;
		++it;
		(*it) >> z;
		cloud.emplace_back(x, y, z);
	}
}