Edit parameters

Figure 3.8: Edit tools; (i) Edit settings, (ii) Focal length, (iii) Rectify, (iv) generate point cloud, (v) Set coordinate reference system

Click the Pen edit button or select Settings->Edit or press keyboard E to edit parameters. This opens the project settings dialog.

There are several experimental parameters that are marked as Advanced. These should in general not be changed.

The most typical parameters to e.g. improve number of matched points or make the process faster are (see detailed description below):

• Feature threshold
• Match threshold
• Search space
• Match window size

Project settings dialog
Set start frame
Focal length
Point cloud rectification
Redo
Coordinate reference system
Dense matching
Change depth of stereo video
Time synchronization

Project settings dialog

Figure 3.9: Settings dialogs for videos and images

• Last pyramid level. To maintain real time performance, it may not be possible to do image matching all the way to the original resolution while the videos are playing. The last pyramid level defines at which image pyramid level image matching stops while running a project. processing will continue to the full resolution when a project is paused, see more here.
• Feature threshold. Edges are extracted in each video frame with sub pixel accuracy. The edge threshold defines the minimum strength of edges. Lowering this threshold will thus create more edges (but in general also more noise).
• Match threshold. The match threshold defines the minimum similarity between matched edges for the match to be accepted and used for sensor modelling. Lowering this threshold will accept more matches (and increase the risk to include more incorrect matches).
• Search space. Search space defines the neighbourhood of pixels to examine for matches. For most parts this is automatically set, but this threshold defines the search space at each pyramid level when a stable sensor orientation has been found. A larger search space is automatically used when the solution is unstable.
• Image registration. A course image registration is performed at project startup and whenever matching is automatically identified as unsuccessful. If there are certain metadata, such as camera position and orientation, then a module for metadata based image registration is automatically selected. If such metadata are missing, then a less accurate translation based image registration is selected. You can select the translation based image registration although there are metadata, but not the opposite.
• Refine sensor model. Leave this on unless you prefer slightly faster processing at the cost of a decreased accuracy.
• Reuse cameras. Leave this unchecked. When checked, it will start estimation of sensor parameters by using the camera parameters of the last stereo pair instead of estimating parameters from scratch, independent of previous parameters. Even when this works it will often degrade the 3D model and its rectification incrementely for each new stereo pair.
• Median variance. To eliminate outliers a variance of unit weight needs to be estimated. The estimated variance will itself be biased by the outliers with the default behaviour. When this is checked, the variance will be computed with a more robust median estimation. However, the effect is in general neglible, so the recommendation is to keep this unchecked.
• Edge focusing. When checked, matched edges at one pyramid level are propagated to the next level. When unchecked, edge pixels in one level will search for successful matches at a previous level. Recommendation is to leave this checked.
• Compute 3D accuracy. When this is checked, the plane and height accuracy of the 3D coordinates are estimated. They are not used internally, but will be exported to ibt files. To avoid unnecessary processing it is recommended to keep this unchecked unless you want to export plane and height accuracies to file.
• Rectify. See the rectify sub section.
• Y parallax threshold. The vertical search space in the epipolar images is based on the accuracy of the last sensor parameter estimation. This value defines how many standard deviations the search space should be. Default is three.
• Z filter, Z low and high thresholds. Extreme high and/or low height values are eliminated from the 3D point cloud when this is checked.
• Percent outliers. If set to any other value than zero, then sensor parameters are estimated with RANSAC. Percent outliers is not a correct naming since you need to set this to values around 25-50Default and recommended is to not rely on RANSAC and instead eliminate outliers integrated with the iterative sensor parameter estimation.
• Feature type. Select if points or edges should be used as matching features. Points as feature type is currently unstable, so the recommendation is to use edges.
• Interpolation. Cubic convolution (default)improves the quality of generated 3D points compared to bilinear interpolation.
• Video stabilization. Individual sensor orientations may be uncertain, but assuming a smooth trajectory we can use previous sensor orientations to stabilize the resulting stereo video. This is thus not used for video stabilization of the video sources.
• Match window. The size of the square window that is used as neighborhood for each image point in template matching. In general, a larger window will result in more robust results, but details may be lost and processing becomes slower.
• Noise filter. Kind of noise filtering applied before edge extraction and matching takes place. Larger filters give less details, but may be more robust. Less edges are created when you increase filter size, so you may want to decrease the edge threshold when you increase filter size. The filter is only applied on internal image for processing, not the the displayed image/frame.
• Point cloud format. Defines point cloud file format when exporting 3D point clouds. PCD is significantly faster to save, but can not handle coordinate reference systems.

Set start frame

Focal length

Point cloud rectification

Redo

The Redo button will regenerate a 3D point cloud from the current images, possibly after you have changes some parameters. To enable a complete refresh, the image registration module is always run when redo is clicked, while it in general is not run while playing or at pausing.

Coordinate reference system

Dense matching

Settings->Dense matching->Settings opens a dialog with settings for generation of a dense point cloud. All settings in this dialog are experimental and not intended to be changed! In particular, the check box Dense matching should be unchecked. Eventually this will generate truely dense point clouds with approximatelly one height estimate per pixel, but this is work in progress.

Change depth of stereo video

Select Settings->Increase offset (Ctrl+Alt+I) or Settings->Increase offset (Ctrl+Alt+D). Take a look at the status bar at the bottom left to see whether frame offset is increased or decreased, since this depends on which frame that is set as left and right.

Time synchronization