Description of the Options
Figure 1. Options dialog with default values |
Loop-closure Search Options
Select for each pose: This option controls how many and which loop-closures are applied for each pose. Possible values are:
- Best (min trace): One loop-closure with the lowest uncertainty will be applied. This is the default and recommended option.
- First: The first loop-closure found will be applied. This will be faster than the 'best' method, however weak (uncertain) loop-closures from the most distant poses are often found first.
- All: All discovered loop-closures will be applied to the SLAM filter. Since each additional loop-closure slows down the SLAM processing and increasingly smaller corrections obtained from applying additional constraints to nearby-poses, using the 'Best' option is recommended instead of this option.
Cross-track only: This option attempts to restrict the search for loop-closures to pairs to images acquired of different mission legs or tracks by using a time threshold. This is useful to speed up the search when you are not interested in matching an image to those acquired at the previous or subsequent few poses (along-track matching).
Min loop time: This option is only valid when 'Cross-track only' has been selected. This defines the minimum time required between when two image pairs were acquired for them to be considered 'cross-track'.
Image Options
Image scale: The images can be down-scaled to reduce computation time at the expense of a reduction in the number of extracted features. We typically use a value of 0.5 corresponding to reducing both the width and height of the image by half, producing an approximately four-fold reduction in processing time. This option can be increased to 1.0 if you wish to attempt to find additional features. Don't use values larger than 1.0, instead try the 'Process images at double scale' option.
Normalise images: If this option is enabled, the images are pre-processed to normalise the pixel intensity values to have a specified mean and standard-deviation by through a linear scaling function. Enabling this option is a good idea if some of the images are overly dark or bright and lack contrast. Since the feature detectors find areas of local contrast, this option is useful to enable a reasonable number of features to be detected in most images using a single default feature detection threshold. I recommend using this option.
Normalised mean: The desired mean pixel value over the whole image. This option is only valid when 'Normalise images' is selected. This value should not affect the number of features found unless the image contains very dark or bright regions where the intensity values are being truncated to 0 or the maximum pixel value. Since there is currently no way to inspect the normalised images, it is recommend that this option is left at the default value of 128 which is half way between the minimum and maximum pixel intensities of 0 and 255.
Normalised standard deviation: The desired standard-deviation of pixel intensities in the image. This option is only valid when 'Normalise images' is selected. Increasing this value will produce more contrast, resulting in more features. While the value of this option has had little experimental testing, the default value of 30 seems fairly reasonable. If the detection of more features is desired, it should be achieved by reducing the feature detection threshold instead of modifying this value.
Feature Options
Feature type: A few feature detectors may be available if supported by the platform being used. SURF is recommend when it is available (on 32-bit computers), otherwise use SIFT. SIFT is slightly better at matching features, however it is significantly slower than SURF. The other feature types are probably only worth trying as a last resort when SIFT and SURF fail to find any matches.
Detection threshold: The feature detection threshold is a magic number tuned for each feature detection algorithm to find a reasonable quantity of features in most images. These thresholds set the minimum level of contrast required locally in an area of the image for the detector to decide a feature is present. Increasing this number will reduce the number of features found, while decreasing the threshold will increase the number of features. Since this option is a threshold on contrast, vibrant images of a coral reef with a higher level of contrast will probably produce more features than washed-out images of texture-less sand. The image normalisation pre-processing option can help produce a reasonable number of features in most images using a set detection threshold.
Process images at double scale: This option, which is only supported by some feature types, causes the feature detection algorithm to search for large-scale features which would normally only be found if the input image was twice as large. This option is designed to improve matching an image to another image acquired at a greater distance. This option is off by default, since image are typically down-scaled, and it is better to increase the 'image scale' to 1.0 before enabling this option.
Modifying the Option Values to Find Additional Loop-Closures
If the default loop-closure search options fail to find any matches in a particular area, it is worth trying with a few different sets of options. Use the following instructions only as a guide. If an optimal method for finding all loop-closures was known, then the process would be automated.
The first thing to do is inspect a few images in the area of interest by clicking on the black dots representing past vehicle poses in the seabed_slam GUI. If the images are completely black or white or are missing, changing the options will not work.
The following is a list of changes to the default values are presented in the order in which in which it is recommend that they are attempted:
- Try using image normalisation if you are not already. This will often help
when some images are too dark.
- If SURF features are used by default, change to SIFT. If you are already
using SIFT do not bother trying the others yet.
- If SIFT fails at the default threshold, try reducing it by half. This will
produce many features, however it will be slow.
- If the changes listed above have failed, try increasing the image scale to
1.0. You probably want to do this first at the default feature detection
threshold, then try again with the reduced threshold. The next option to try
is 'Process images at double scale'.
- If nothing has worked yet, you could try the other feature types (e.g. CV-SURF), however at this point you are probably just wasting your time. If SIFT has not worked, then the other feature types probably will not either.