Release Notes for HALCON 20.05 Progress

• The following licensing issues are known:
  • After starting applications based on HALCON 20.05 (or later), the dongle
    must be removed and inserted again before starting applications based on
    HALCON 19.11 (or earlier).
    This problem can be avoided by installing the CodeMeter Runtime from
    WIBU SYSTEMS AG.
    The CodeMeter Runtime is not shipped with HALCON and must be downloaded from
    https://www.wibu.com in the download area.
  • On Arm platforms (aarch64 architecture) with GLIBC 2.21, it is no longer
    possible to run both 64-bit and 32-bit HALCON processes in parallel because
    they might block each other.
  More information.

• Since find_box_3d is now independent of a viewpoint and the position of the input point cloud in space, the procedure debug_find_box_3d has been adapted accordingly. Furthermore, an error is raised in case that the input GenParam contains a key 'viewpoint'. The returned GrippingPose is set according to the XYZ-mapping.
  If the input point cloud of find_box_3d contains normals, they are not used anymore. More information.
• In rare cases, find_shape_model (as well as the scaled and aniso variants) might return different results. More information.
• Because of a bug fix regarding the calculation of the mean average precision (mean_ap) measure during the evaluation of deep-learning-based object detection, the deep-learning-based object detection evaluation results might be different. To obtain the same results for 'mean_ap' as before, the evaluation parameter 'interpolate_pr_curves' can be set to false (default: true). More information.
• Due to changes to the calculation of the anomaly score histogram in evaluate_dl_model (in order to solve a display issue), the following parameter keys have become obsolete and hence unavailable for the procedure's generic parameter dictionary GenParam:
  • 'anomaly_num_bins'
  • 'anomaly_min_value'
  • 'anomaly_max_value'
  More information.
• For deep-learning-based object detection models, get_dl_model_param now returns a tuple of values that is as long as the number of classes for the parameter "class_weights" in any case. More information.
• reconstruct_surface_stereo now uses the default values for the parameters 'binocular_mg_initial_guess' and 'binocular_mg_initial_level' if no values have been set with set_stereo_model_param. This might lead to slightly different reconstruction results, since previously both parameters have been estimated based on the set bounding box. More information.
• The default value for the texture inspection model parameter 'gmm_preprocessing' has been changed from 'none' to 'normalization'. When using gray scale images, identical results compared to previous HALCON versions can be obtained by calling set_texture_inspection_model_param(TextureInspectionModel, 'gmm_preprocessing', 'none') before executing train_texture_inspection_model(). More information.
• In previous versions, train_dl_model_batch expected the parameter 'DLSampleBatch' as single dictionary in the C++ and .NET interface. While the operator now expects a tuple of dictionaries, with the new signature it is still possible to use a single dictionary for 'DLSampleBatch'. More information.
• For HALCON 20.05 and newer, only NVIDIA Cards with compute capability 3.5 are supported. More information.

Code Export

GUI

IDE

Language

Code Export

GUI

Help

IDE

Procedures

Miscellaneous

New HDevelop Example Programs

• The just-in-time compiler of the HDevEngine is now able to compile procedures that call dev_clear_obj on a global variable.
• The C# and VB HDevEngine examples for executing programs and procedures now use a background thread for execution. This provides a better experience when these examples are adapted to execute own code that makes use of interactive drawing objects.
• JIT compilation of procedures is now supported with the standard Arm architectures aarch64-linux and armv7a-linux.
  The size of the HDevEngine for Arm architectures increased to approximately 30 MB.

• For some special cases, add_image used an MMX implementation that wrote results outside the actual region. This problem has been fixed.
  As a consequence of the fix, add_image is now up to 50% faster for images of type 'byte', 'uint2', and 'int2' for Mult = 1.0 and Add = 0.0.
  For Mult = 0.5 and Add = 0.0 a performance improvement of up to 200% can be expected for images of type 'byte', up to 250% for 'int2' images, and up to 120% for 'uint2' images.
• The following operators have been sped up and parallelized internally when operating on many points:
  affine_trans_pixel, affine_trans_point_2d, affine_trans_point_3d, projective_trans_pixel, projective_trans_point_2d, projective_trans_point_3d, projective_trans_hom_point_3d, project_hom_point_hom_mat3d, project_3d_point, project_point_hom_mat3d, rigid_trans_object_model_3d, affine_trans_object_model_3d, projective_trans_object_model_3d, and get_line_of_sight.
  
• compare_variation_model and compare_ext_variation_model for all supported image object types ('byte', 'int2', and 'uint2') have been optimized to use the Arm Neon instruction set extension. On a Cortex A73 processor, the operator is now 220% faster for 'byte' images.
• compare_variation_model and compare_ext_variation_model are now faster for all supported image types ('byte', 'int2', and 'uint2') on Intel compatible processors that support AVX2. In particular, the following speedups can be expected compared to the standard C implementation:
  

  Image Type	Operator	Mode	Available Instruction Set	Speedup
  byte	compare_variation_model	-	AVX2	up to 475%
  int2		-		up to 625%
  uint2		-		up to 550%
  byte	compare_ext_variation_model	'absolute'		up to 400%
  int2				up to 550%
  uint2				up to 500%
  byte		'light', 'dark'		up to 700%
  int2				up to 250%
  uint2				up to 250%
  byte		'light_dark'		up to 675%
  int2				up to 550%
  uint2				up to 380%

   
• derivate_gauss for all image types are now supported on computers that support the AVX512F instruction set. For 'byte' images, the operator is 30% faster than the implementation in AVX2.
• The SIMD implementations for dyn_threshold returned wrong results in some cases if the values for the parameter 'Offset' exceeded the maximum for the used datatype (i.e., >= 256 for 'byte' and >= 65536 for 'uint2' images). This problem has been fixed.
  Due to a revised implementation, dyn_threshold is now faster if the parameter 'Offset' is set to values smaller than zero for 'light' and 'dark'. In particular, the performance improvement is up to 20% for 'byte' images and 50% for 'uint2' images.
• There is a significant speedup for the following deep learning procedures, especially for long input or output tuples:
  • find_dl_samples
  • split_dl_dataset
  • gen_dl_samples
  This also affects other deep learning procedures that use the procedures mentioned above.
• list_files is now faster on Unix systems by a factor of at least 2.
• mirror_image with 'byte' images, Mode='col', and image widths that are a multiple of 4 is now faster on Intel x86 compatible machines. The speedup is up to 270%.
  Furthermore, mirror_image is now faster with 'byte' images, Mode='col' on Intel compatible processors that support AVX2 or SSSE3. In particular, the speedup is up to 450%.
  
• projective_trans_image crashed in rare cases if the interpolation was set to 'nearest_neighbor'. This problem has been fixed.
  The fix led to a speedup of projective_trans_image and projective_trans_image_size. The acceleration for all supported image types is up to 8% for processors that support AVX2 and up to 20% if AVX2 is not supported. The acceleration is only valid if 'int_zooming' is set to 'true', which is the default.
  
• The operator read_object_model_3d is now able to read binary PLY files that contain previously unsupported elements and properties. If possible, such items are converted into extended attributes while reading. Additionally, reading PLY files is now up to 380% faster when reading PLY ASCII files and up to 72% faster when reading binary PLY files that contain faces.
• set_dl_model_param has been extended to set the parameter 'fuse_bn_relu'.
  If this parameter is set to 'true', each of the batch normalization and activation layers in a deep learning network will be fused. Currently, only ReLU activation is supported. As a consequence, the runtime for the inference and for a training step can be reduced. This function has been adapted for the following delivered networks:
  • initial_dl_anomaly_medium.hdl: Speedup up to 12 % for inference and up to 15 % for a training step.
  • initial_dl_anomaly_large.hdl: Speedup up to 13 % for inference and up to 1 % for a training step.
  • pretrained_dl_classifier_resnet50.hdl: Speedup up to 15 % for inference and up to 8 % for a training step.
  • pretrained_dl_segmentation_compact.hdl: Speedup up to 15 % for inference and up to 7 % for a training step.
  • pretrained_dl_segmentation_enhanced.hdl: Speedup up to 19 % for inference and up to 14 % for a training step.
  
  Note that this change affects the compatibility. Read more.
• threshold and dyn_threshold have been optimized for the Arm Neon instruction set extension for 'byte' and 'uint2' image types, for both 32 and 64 bit Arm systems. For the threshold operator, operations on float images have been optimized as well. On Cortex A73 processors, for 'byte' images threshold operations are up to 200% faster, while dyn_threshold operations are up to 180-250% faster, depending on the mode.
• train_class_gmm is now significantly faster when training a classifier with a single class. Furthermore, train_class_gmm is now also parallelized on an internal data level when training a classifier with a single class.
  The achievable speedup depends on the number of available threads and the selected covariance matrix type.
  The following speedups can be expected for classifiers with high-dimensional training data (n=25) and a large amount of training data (>100.000) with the following exemplary configurations:
  

  Number threads	Speedup for covariance 'diag'	Speedup for covariance 'full'
  1	139%	306%
  2	261%	535%
  4	341%	957%
  8	375%	1530%
  16	463%	2346%

  As a result, train_texture_inspection_model is now also significantly faster when executed either in parallel or with a single thread.
• train_dl_model_anomaly_dataset has been improved and is now faster. A speedup up to 900% and 1500% can be expected for CPU and GPU training, respectively.
• The thirdparty Arm Compute Library (ACL) has been updated from version 18.11 to 19.08. The BatchNorm layer was included. As a result, the delivered deep learning segmentation networks and detection networks can be accelerated by up to 5 %.
• Reading big datasets was rather slow. This problem has been fixed.
• The inference speed of the anomaly detection models initial_dl_anomaly_medium.hdl and initial_dl_anomaly_large.hdl has been improved by up to 50% on GPUs and up to 30% on CPUs.

3D

Bar Code

Calibration

Classification

Data Code

Deep Learning

File

Filter

Graphics

Inspection

Matching

OCR

Parallelization

Segmentation

System

Tuple

3D

Bar Code

Calibration

Classification

Data Code

Deep Learning

File

Filter

Graphics

Inspection

Images

Matching

Matrix

Miscellaneous

OCR

Region

Segmentation

Tuple

XLD

• The HALCON/.NET interface has been extended with support for .NET Core. This support comes in the form of two new packages that are available online via NuGet:
  • MVTec.HalconDotNet
    • This package targets .NET Standard 2.0 and provides full access to the HALCON library as well as HDevEngine. It can be used on all architectures supported by HALCON.
  • MVTec.HalconDotNet-Windows
    • This package targets .NET Core 3.1 and additionally includes Windows Forms and WPF controls for integrating HALCON windows into GUI applications.
  Note that for technical reasons the package versioning deviates from standard HALCON versioning. For HALCON 20.05 Progress, the matching package version is 20050.0.0. For more information please consult the Programmer's Guide. The documentation of the HALCON/.NET interface has been extended with information on using .NET Core.
  The minimal new C# example 'ConsoleApp' demonstrates the use of .NET Core under Windows and Linux. In addition, the C# examples 'Matching' and 'MatchingWPF' have been extended with build configurations for .NET Core under Windows. Instructions for creating a minimal console-based HDevEngine example can be found in the documentation.

• The C# and VB HDevEngine examples for executing programs and procedures now use a background thread for execution. This provides a better experience when these examples are adapted to execute own code that makes use of interactive drawing objects.

• The example c#/DrawingObjects only showed a part of the HALCON window on some systems. This problem has been fixed.

• For the procedure HIncrRL, which increases the current runlength number for memory allocation, a dedicated documentation is provided in the Extensions Package Programmers Manual.

• The list of hbench reference values (hbench -reference) has been extended by benchmark values for the processors AMD Ryzen™ 9 3950X and Intel® Core™ i9-10900X.

• The MVTec GigE Vision Streaming Filter version 2.2.8.3 that is part of the Windows installer fixes two problems:
  When using '[Stream]EventTransferEnd' on multiple NICs, the filter driver sporadically stopped to deliver events. Further, get_framegrabber_param(..., 'num_buffers_underrun', ...) was only available with the socket driver up to now.

• The new HDevelop example programs
  • hdevelop/Image/Acquisition/gstreamer_2pipelines.hdev
  • hdevelop/Image/Acquisition/gstreamer_acquisition_events.hdev
  • hdevelop/Image/Acquisition/gstreamer_nvidia_arguscamera_simple.hdev
  • hdevelop/Image/Acquisition/gstreamer_simple.hdev
  show the usage of the GStreamer interface.