STAIR Vision Library Known Bugs

Bugs marked with have not yet been resolved.

Version 2.x

Bugs are now managed through

sourceforge.

Version 2.0

svlFeatureWhitener does not handle case of constant feature value correctly. (Fixed in build 1770)
Memory leak in svlBoostedClassifier::train(). Reported by Steve. (Fixed in build 1774)
Bug in svlCodebook constructor. Reported by Olga. (Fixed in build 1775)
Function svlClipRect has inconsistent behavior. Change to always clip rectangle to fit into image (instead of sometimes shifting it). Reported by Olga. (Fixed in build 1775)
Missing shuffleDirectory.pl script. Reported by Paul. (Fixed in build 1776)
Setting svlImageLoader and svlTrainingDatasetBuilder through an XML file causes infinite recursion. Reported by Blake. (Fixed in build 1779)
Filenames from the command line for object detection applications need to have the absolute path specified. Reported by Olga. (Fixed in build 1782)
The -cv option of trainObjectDetector erroneously reads two additional arguments (four in total). Workaround for version 2.0 is to add two dummy arguments. Reported by Olga. (Fixed in build 1786)
The -i flag of trainObjectDetector causes a segmentation fault. Reported by Olga. (Fixed in build 1786)
svlImageSegmentation::removeSmallSegments() can sometimes introduce disconnected components. (Fixed in build 1787)
Incorrect argument to loadFrame() in svlImageLoader::getFrameFromFiles(). Reported by Tianshi. (Fixed in build 1790)
Hello World example does not compile against version 2.0. (Fixed in build 1803)
Incorrect function signature for vectorized svlClassifier::getClassScores(). Reported by Paul. (Fixed in build 1809)
Memory access bug in tryThresholds() in svlPRcurve.cpp. Reported by Tianshi. (Fixed in build 1997)
Boosted classifiers trained under Windows fail to load. Reported by Tianshi. (Fixed in build 1820)
svlDisjointSets join fails if both sets have the same ID. Reported by Steve. (Fixed in build 1821)
classifyImages strips multiple extensions off the image filename when writing object files. Reported by Tianshi. (Fixed in build 1843)
svlPatchDictionary patch response calculation uses width instead of widthStep to move to the next row. This could cause errors for images with width not word-aligned. (Fixed in build 1848/1863)
(Linux) make.local variables do not work if there is a comment on the same line. Reported by Adam. (Fixed in build 1870)
Install script (external/install.sh) puts wxWidgets in the wrong place under Linux. (Fixed in build 1870)
Segfaults when extracting random negatives from images with certain parameter settings. Reported by Jason Ma. (Fixed in build 1875/1877)
svlClassifierFactory fails to load boosted classifiers if saved with a .xml extension. (Fixed in release 2.1)
svlLogger log file overwrite flag is inverted. (Fixed in build 1918)
svlBoostedClassifier training fails intermittently with certain large inputs. This is thought to be an OpenCV issue.
Memory leak in buildTrainingDataset when using -bestPosWindow option. (Fixed in build 1952)
svlClassifier fails to build against Eigen2-2.0.2. (Fixed in build 1960)
svlTrainingDatasetBuilder does not work properly with more than one channel of input. (Fixed in build 1991)
svlTrainingDatasetBuilder fails if creatPosDir is false. (Fixed in build 1984)
Dynamic libraries fails to build on Linux without Cuda kernels. (Fixed in build 2009)

Version 1.1

log() bug in implementation of [Sontag, UAI 2008] MAP inference in svlMessagePassing. (Introduced in build 1306. Fixed in build 1509)
mexFactorGraphInference sometimes crashes when given uniform factors. (Fixed in build 1514)
entropy() function doesn't handle zero probabilities. (Fixed in build 1549)
SVL_MP_ASYNCLOGMAXPROD_LAZY message passing may return incorrect results on graphs with multiple cliques over the same variables.
svlPatchDictionary does not do a deep copy causing errors for multi-threading. Reported by Ian and Olga.(Fixed in build 1583)
median in svlStatsUtils produces incorrect results. Reported by Paul. (Fixed in build 1677).

Version 1.0

svlCameraIntrinsics::undistort() fails when given images with different depths. (Fixed in build 1194)
GEMPLP graphical models inference produced incorrect results when given empty factors. (Fixed in build 1231)
svlGaussian::evaluate() returned garbage. (Introduced in build 1233. Fixed in build 1266)
svlSegImage set default name from image filename incorrectly. (Fixed in build 1269)
GUI applications (linked against wxWidgets) do not support unicode and fail to build on some systems.
Compilation errors on (Linux) systems with g++ 4.3 due to change of header inclusions. Reported by Michael Levin. (Fixed in build 1352)
Some mex routines include links to unneeded libraries (in particular libcv.so).
svlPatchDictionary can produce NaNs when run on regions with uniform intensity. Reported by Charles DuHadway. (Fixed in build 1324)
Bug in estimatePointNormals and estimatePlane. Reported by Shu Tsukiji. (Fixed in build 1346)

Version 0.9

regionLabeler crashes under Windows when loading definitions file. Problem is related to speedbar buttons. (Fixed in build 876)
scoreDetections gives incorrect precision-recall curve when there are no detections. (Fixed in version 2.0)
scoreDetections computes the wrong precision-recall curve even when there are detections. (Fixed in build 872)
dirent can load files in different orders depending on the machine. This causes runTests.pl to sometimes report a regression test failure even through the code is working correctly. (Fixed in build 863)
imageSequenceLabeler displays images upsidedown when undistort is turned off on Windows. (Fixed in build 852)
buildPatchDictionary has a memory leak. (Fixed in build 863)
svlConfusionMatrix produces incorrect results for function totalSum() when confusion matrix is non-square. (Fixed in build 867)
svlOptimizer aborts when trying to solve problems with less than 5 variables. This causes problems for code that use the optimizer, for example, when learning multi-class logistic classifiers with less than 5 parameters. (Fixed in build 929)
svlOptimizer requires an explicit call to initialize() to initialize parameters. This caused problems with CRF learning. (Fixed in build 1026)
make clean does not properly remove all object files. (Fixed in build 1036)