Difference between revisions of "BoofCV Change Log"

From BoofCV
Jump to navigationJump to search
m
Line 2: Line 2:


Date Format: year/month/day
Date Format: year/month/day
---------------------------------------------
Date    : 2012/11/6
Version : Alpha 0.11
* Association Score
** Normalized Cross Correlation (NCC) has been fixed (yet again) in ScoreAssociateNccFeature where the abs() was incorretly being taken.
** Fixed bug in NccFeature.setTo()
* FactoryAssociation
** Added NCC to defaultScore()
* Template Matching
** Added NCC template matching algorithm
** TemplateScoreType: ERROR_SQUARED renamed to SUM_DIFF_SQ
* ModelSet
** Original index of inlier set can be requested now
** Removed the ability to generate multiple hypotheses.  Was not needed after all.
** RANSAC score has been deleted and RANSAC simplified
* RectifyImageOps
** Cleaned up function names and improved documentation
* LensDistortionOps
** Added transformRadialToPixel_F64()
* Reorganized functions in geometric vision package
** Created MultiViewOps and PerspetiveOps
** Refactored/deleted other related Util classes
** Enums are now used to select which algorithm is returned in the Factory in many cases
** New interfaces for each problem and improved JavaDoc
* Moved AssociatedPair and PointPositionPair to boofcv.struct.geo package
* Trifocal Tensor
** Added related functions to MultiViewOps
** Added linear estimator for 7+ points
* PnP Problem
** Added two P3P solutions
** Analytic Jacobian for non-linear refinement
** Refactored class names for better consistency
* Fixed bugs in StereoSparse3D
* Fixed bug in DistortImageOps.boundBox_F32()
* Added ImageGridPanel to display a set of images as a grid
* ModelMatcher added getMinimumSize()
* Renamed PointPosePair to Point2D3D since the old name was misleading
* Visual Odometry
** Added stereo camera based algorithm that uses PnP for pose estimation
** Removed broken mono and stereo algorithms
* DetectAssociateTracker
** Changed from an abstract class into a regular class
** Deleted SURF,BRIEF,NCC specific implementations since they are no longer needed
* AssociatedPair
** renamed key and curr points to p1 and p2
** new names are more generic and consistent
* DetectAssociateTracker
** Changed internal design a bit and fixed bugs
** Performs well in structure from motion applications (2D and 3D)
* ImagePointTracker accessors changed so that a storage list is requested
** Makes usage easier since now its impossible to drop a track while reading from the main list.
** Updated implementations so that the internal list is not returned
** Removed addTrack() since only KLT supported that functionality, hence not generic.
* ConvertBufferedImage
** Fixed bug where INDEXED images where not being converted correctly


---------------------------------------------
---------------------------------------------

Revision as of 18:23, 6 November 2012

Change Log

Date Format: year/month/day


Date  : 2012/11/6 Version : Alpha 0.11

  • Association Score
    • Normalized Cross Correlation (NCC) has been fixed (yet again) in ScoreAssociateNccFeature where the abs() was incorretly being taken.
    • Fixed bug in NccFeature.setTo()
  • FactoryAssociation
    • Added NCC to defaultScore()
  • Template Matching
    • Added NCC template matching algorithm
    • TemplateScoreType: ERROR_SQUARED renamed to SUM_DIFF_SQ
  • ModelSet
    • Original index of inlier set can be requested now
    • Removed the ability to generate multiple hypotheses. Was not needed after all.
    • RANSAC score has been deleted and RANSAC simplified
  • RectifyImageOps
    • Cleaned up function names and improved documentation
  • LensDistortionOps
    • Added transformRadialToPixel_F64()
  • Reorganized functions in geometric vision package
    • Created MultiViewOps and PerspetiveOps
    • Refactored/deleted other related Util classes
    • Enums are now used to select which algorithm is returned in the Factory in many cases
    • New interfaces for each problem and improved JavaDoc
  • Moved AssociatedPair and PointPositionPair to boofcv.struct.geo package
  • Trifocal Tensor
    • Added related functions to MultiViewOps
    • Added linear estimator for 7+ points
  • PnP Problem
    • Added two P3P solutions
    • Analytic Jacobian for non-linear refinement
    • Refactored class names for better consistency
  • Fixed bugs in StereoSparse3D
  • Fixed bug in DistortImageOps.boundBox_F32()
  • Added ImageGridPanel to display a set of images as a grid
  • ModelMatcher added getMinimumSize()
  • Renamed PointPosePair to Point2D3D since the old name was misleading
  • Visual Odometry
    • Added stereo camera based algorithm that uses PnP for pose estimation
    • Removed broken mono and stereo algorithms
  • DetectAssociateTracker
    • Changed from an abstract class into a regular class
    • Deleted SURF,BRIEF,NCC specific implementations since they are no longer needed
  • AssociatedPair
    • renamed key and curr points to p1 and p2
    • new names are more generic and consistent
  • DetectAssociateTracker
    • Changed internal design a bit and fixed bugs
    • Performs well in structure from motion applications (2D and 3D)
  • ImagePointTracker accessors changed so that a storage list is requested
    • Makes usage easier since now its impossible to drop a track while reading from the main list.
    • Updated implementations so that the internal list is not returned
    • Removed addTrack() since only KLT supported that functionality, hence not generic.
  • ConvertBufferedImage
    • Fixed bug where INDEXED images where not being converted correctly

Date  : 2012/09/18 Version : Alpha 0.10

  • Association Score
    • Pushed implementations of ScoreAssociation into abst package
    • Put most association distance algorithms into DescriptorDistance
    • Added hamming lookup table
    • Brief now uses a look up table. About a 7x speed improvement.
    • FactoryAssociate now can create instances of ScoreAssociate.
    • Fixed incorrect NCC equation. Thanks Christian Kramer.
  • DescribeRegionPoint has been redesigned
    • Now more generic and can work with any feature descriptor
    • Separate function for checking to see if feature is inside bounds
  • TupleDesc
    • changed to an interface
    • functions added to provide some generic support
  • Brief descriptor renamed to TupleDesc_B
    • Acts as a generic binary feature descriptor
  • Descriptors
    • BRIEF and BRIEF-SO now can process features along the image border. Bit set to false in that case.
  • Visualization Apps
    • VisualizeAssociationScoreApp has score type in a separate panel and can support binary descriptors
  • InterestPointDetector
    • WrapFHtoInterestPoint now returns the correct detection radius
    • Fixed non-compliant implementations with incorrectly handled unsupported features. Thanks jasonnn for submitting this bug
  • FAST Feature Detector
    • Original implementation was buggy and did not produce same results as original author's code
    • Rewrote from scratch and replicates original author results
  • GeneralFeatureDetector moved from extract to interest package
  • Fundamental/Essential Matrix
    • 7-point algorithm now returns all solutions and the algorithm was tweaked
    • Nister 5-pt Essential Matrix
    • Abstractions which can handle multiple solutions
    • Stability evaluator for epipolar matrices
    • Improved unit tests
  • Polynomials
    • Various ways to find and polish polynomial roots (EVD and Sturm)
    • New data type and basic operations
  • Added ExampleStereoTwoViewsOneCamera
    • Demonstrates how to perform stereo image processing using a single camera and two views
  • LensDistortionOps
    • Added normalized to distorted pixel transform
    • Tweaked some of the function names
  • SparseIntegralSample_I32.isInBounds() did not return that correct value along the right and bottom image border
  • Functions inside of GThresholdImageOps did not return the number of blobs found like it should
    • Thanks jasonnn
  • NonMaxBlockStrict
    • Fixed bug where it assumed intensity values were all positive
  • TemplateMatching
    • Added code for template matching
    • Added example code to demonstrate its use
  • Added support for alpha channel in ConvertBufferedImage
  • FactoryPointSequentialTracker doesn't create trackers with FAST detector instead uses ShiTomasi to improve stability

Date  : 2012/7/22 Version : Alpha 0.9

  • Renamed factories
    • FactoryBlobDetector -> FactoryDetectPoint
    • FactoryCornerDetector -> FactoryDetectPoint
  • Renamed KLT corner to Shi Tomasi corner, which is more standard
  • Fixed bug various point detectors where if the input image was resized the edge intensity might not be zero when it should be
  • Changed some corner detection algorithms to take the intensity image in as an input.
  • Added new Pixel to Normalized transform
  • Created interface for Stereo visual odometry
  • Fixed bug in unsafe_set inside of ImageInt8
  • Added unsafe_set() and unsafe_get() to ImageFloat32, ImageFloat64, ImageSInt64
  • Fixed bug in ImageSInt64.set() where the value was cast into a short
  • Changed transform* functions inside of DistortImageOps so that you can specify how outside image pixels are handled.
  • Fixed bug in ImageHistogramPanel - reported by Ray
    • Added support for integer images
  • Integration
    • Created ant build scripts
    • Updated Xuggler integration code
    • Added Android
  • Generalized SimpleImageSequence so that it does not reference BufferedImage and will play nicer with Android
  • Fixed bug in ImplImageDistort* where for 16bit and 32 bit images values were being type casted into bytes
  • Changed UnconstrainedMinimization so that the minimum function value is specified in setFunction()
  • Added ExampleContours
    • Thanks cspnanda for pointing out the lack of documentation
  • New non-maximum suppression algorithm
    • About the same for 3x3 regions, much faster for larger regions
    • Tweaked related API and improved unit tests
  • Fixed bug in WrapDescrivePixelRegion where the descriptor is not properly configured for integer images

Date  : 2012/05/13 Version : Alpha 0.8

IMPORTANT Discard old calibration files and calibrate again with latest code!

  • Added processing monitor to CalibrateStereoPlanarGuiApp
  • Fixed bug in Zhang99 calibration where Rodrigues coordinates were being parameterized incorrectly
    • Calibration results would be less accurate, but have a lower residual due to the increased degree of freedom
  • Added analytic Jacobian to Zhang99
    • Decided not to use it since it produces same estimation parameters as numerical.
  • Added functions for creating detectors in FactoryPlanarCalibrationTarget
    • changed code to reference those functions
  • Stereo Disparity Image
    • Dense and Sparse algorithms
    • Rectangular region winner take all (WTA)
      • Validation: right to left, max error, texture
      • Precision: Pixel and Subpixel
    • Algorithms:
      • Basic rectangular region
      • Five region adaptive: Hirschmuller, et. al. "Real-Time Correlation-Based Stereo Vision with Reduced Border Errors," 2002
  • IntrinsicParameters
    • Added flipY to parameter list since it is vital information for many calculations.
  • RectifyImageOps
    • Fixed bug in functions that adjusted the rectification for viewing. If left handed the coordinate systems got messed up
  • Changed allInside() so that the annoying single pixel black region is gone
    • RectifyImageOps and LensDistortOps
  • FactoryCornerDetector
    • Added flag for weighted KLT and Harris
  • Added weighted corners to list of features in some evaluation apps
  • Examples:
    • ExampleCalibrateMonocularPoints
    • ExampleDetectCalibrationPoints
    • ExampleStereoDisparity
  • ImageTestingOps
    • Flip image vertical
  • Fast point cloud viewer

Date  : 2012/04/22

Version : Alpha 0.7

  • Redesigned how evaluation apps/applets handle data input and configuration
  • Removed the epipolar directory
    • More than just stereo algorithms in there
  • Camera calibration outputs XML file containing intrinsic parameters
    • Java XML serialization
  • For square calibration grids, added sub-pixel refinement using canny edge
  • Fixed a bug in camera calibration where it transposed the grid instead of rotating it if the number of rows/columns needed to be swapped.
  • 3D Vision Algorithms
    • Estimate translation given rotation and two observations
    • Refine pose estimate given set of 3D points and observations
    • Interface for n-view triangulation
    • Bundle adjustment for calibrated cameras
      • Analytic Jacobian
      • A bit slow right now
  • Changed behavior of GeneralFeatureDetector
    • Supports sub-region extraction. Forces features to be more spread out
    • Max N now is the total number of features, including excludes
  • Improved speed of Levenberg algorithm by taking advantage of symmetry
  • UtilEpipolar
    • Added canonicalCamera() and decomposeCameraMatrix()
  • Created the following Util and Ops classes
    • UtilIntrinsic
    • LensDistortionOps
    • RectifyImageOps
  • Stereo Rectification
    • Calibrated with known baseline
    • Uncalibrated, from using a fundamental matrix.
    • Adjust rectification to maximize usefull viewing area
  • New Examples
    • Estimate fundamental matrix
    • Rectify calibrated
    • Rectify fundamental matrix
    • Remove lens distortion
  • The feature score was not being set in WrapAssociateGreedy
    • Thanks Nir Amar
    • Also changed AssociateSurfBasic for a similar reason
  • Fixed problem in FundamentalResidualSampson where the denominator was being squared
  • Fixed problem in Se3 Essential generator
  • Changed positive depth constraint class

Date  : 2012/03/01

Version : Alpha 0.6

  • Added convolution related code for ImageSInt32
  • Added weighted KLT and Harris corner detectors
  • Refactored factory names and other classes to make them easier to find and more consistent
  • Planar calibration grid detectors
    • Chess board pattern
    • Square grid pattern
  • Generic code for detecting squares in binary images
    • Inside of calibration module
  • Application for calibration from planar calibration targets
    • GUI and commandline
  • ImageDistort
    • Fixed bug where border.setImage() was not being called
  • ImageBorder
    • Added generalized get(x,y)
  • Refactored FactoryImageBorder into two factories
  • Fixed bug in several derivative operators where the passed in border was being ignored and an extended border being used instead
  • Change ImplImageDistort* to use interpolate.get_unsafe() to speed it up by a large margin
  • Moved ImageDistort from boofcv.alg.distort to boofcv.struct.distort
  • Added PointTransform_F32
  • Added both forward and reverse radial distortion PixelTransform_F32 implementations
  • ModelMatcher classes can now create multiple model hypotheses from a single set of points.
  • Changed ImagePointTracker to use PointTrack instead of AssociatedPair
    • Created KeyFramePointTracker to track features and create AssociatedPairs
  • Added FactoryEpipolar
  • Added FactoryTriangulate
  • Added non-linear refinement for:
    • Fundamental/Essential matrix
    • Homography matrix
    • Triangulation
  • Added 2-view and N-View triangulation algorithms
    • DLT
    • Closest point from two lines
  • Added EPnP from Lepetit 2009 paper
    • Validation against author's code is pending
  • Added color processing to several Apps
    • EvaluateInterpolateEnlargeApp
    • ShowImageBlurApp
  • Added SimpleNumberSequenceReader for reading small sequences of numbers
  • Added lens distortion app
  • Found sudbtle bug in Harris corner detector.
    • Surprisingly doesn't seem to change output.

Date  : 2012/02/01

Version : Alpha 0.5

  • Improved SURF performance
    • Fixed various bugs and changed algorithm to improve performance
      • Both speed and stibility have been improved
    • Now is the best performing SURF implementation in the benchmark
  • ImplOrientationSlidingWindowIntegral has been replaced with a faster and more accurate algorithm
  • Improved integral image support
  • Improved sparse gradient support
  • Refactored sparse image operator code
    • Added SparseImageSample interface
  • Added/Improved unconstrained non-linear optimization code
    • Quasi-Newton line search
    • Levenberg-Marquardt
  • Geometric vision and calibration code which uses non-linear optimization now works correectly
  • Started converting internal micro benchmarks to use Caliper

Date  : 2012/01/06

Version : Alpha 0.4

  • Changed behavior of ConvertBufferedImage so that it won't automatically swap bands in MultiSpecitral image
    • orderBandsIntoRGB() has been provided to do that manually if needed
  • All automatic convert to/from BGR and been removed from ConvertRaster
  • Fixed SURF related issue
    • Image bounds check was being incorrectly performed for integral image orientation estimation algorithms (Thanks Oleg Chernoguz for pointing this out)
    • FactoryOrientationAlgs.sliding_ii() was creating an average orienation alg for I32 images
    • Some coordinates were integer instead of double
  • Added comments to several factories and other classes
  • Added ant script for examples and readme.txt explaining how to use it

Date  : 2012/01/02

Version : Alpha 0.3

  • Camera Calibration
    • Planar grid (Zhang 98)
    • Camera calibration matrix from homographies
    • Linear radial distortion
    • Decomposition of homography for calibration
    • Nonlinear optimization of all parameters
  • Calibration Grid Detection
    • Detects corners up to "pixel accuracy", in reality its probably worse than that
    • Sub-pixel accuracy code exists but needs more work
  • Added min and max values to ImageTestingOps.addGaussian()
  • Tweaked factories related to wavelet denoising so that they take in an image type.
  • Removed cyclical dependency between modules by creating a visualize modules
  • Added to BinaryImageOps
    • label to clusters
    • clusters to binary
  • Renamed ImageBase to ImageSingleBand
  • Created a new ImageBase
  • Flushed out MultSpectral image type more and extended ImageBase
    • Created functions to convert to and from BufferedImage
  • Changed behavior of several functions in GeneralizedImageOps to include MultiSpectral images
    • refactor createImage -> createSingleBand
  • Updated ant scripts
    • Common template is now used for the different modules
    • Fixed unit test and make it work across the whole project
  • Added crop to ImageDistort
    • If cropping is used the only part of the destimation image is processed
  • Created calibration package to hold calibration related code
  • Moved jars not part of the primary build into boofcv/lib/testing
  • Started converting runtime benchmark code over to Caliper

Date  : 2011/12/01

Version : Alpha 0.2

  • Added integer polynomial interpolation
    • Cleaned up polynomial interpolation code
  • Fixed bug in PixelMath.bound() dealing with unsigned images
  • Converting to a BufferedImage now marks it as modified
  • Added min and max pixel values to ImageTypeInfo
  • Added ComplexMath
  • Added PolynomialSolver
  • Added process(image) to PointSequentialTracker and renamed it to ImagePointTracker
  • Added the following 3D vision code:
    • Linear 8-point essential/fundamental matrix
    • Linear 7-point essential/fundamental matrix
    • Linear 4-point homography
    • Extract camera motion from essential
    • Extract camera and plane normal from homography
    • Triangulate: Linear pixel depth
    • Projective N Point (PNP): Linear N >= 6
  • Added non-linear optimizer based upon Levenberg-Marquardt
  • Added pixel region based descriptors
    • Raw pixel values
    • Normalized Cross Correlation (NCC)
  • Added pixel region based trackers
  • Added Sum of Absolute Difference (SAD) score for feature association with a few different descriptor types.

Date  : 2011/11/01

Version : Alpha 0.1

  • Initial public release