Show EOL distros: 

Package Summary

This package contains nodes which make up the Passive-Scene-Recognition interface to Implicit Shape Model (ISM) trees. The Active-Scene-Recognition interface to ISM trees is located in asr_recognizer_prediction_ism, instead. A short outline of the functionalities provided by the nodes are: 1. Recording of scenes 2. Training of an ISM tree (Implicit shape model tree) 3. Recognition of scenes 4. Visualization of ISM (tree) data etc.

  • Maintainer: Meißner Pascal <asr-ros AT lists.kit DOT edu>
  • Author: Borella Jocelyn, Hanselmann Fabian, Heller Florian, Heizmann Heinrich, Kübler Marcel, Mehlhaus Jonas, Meißner Pascal, Qattan Mohamad, Reckling Reno, Stroh Daniel
  • License: BSD
  • Source: git https://github.com/asr-ros/asr_ism.git (branch: master)

Package Summary

This package contains nodes which make up the Passive-Scene-Recognition interface to Implicit Shape Model (ISM) trees. The Active-Scene-Recognition interface to ISM trees is located in asr_recognizer_prediction_ism, instead. A short outline of the functionalities provided by the nodes are: 1. Recording of scenes 2. Training of an ISM tree (Implicit shape model tree) 3. Recognition of scenes 4. Visualization of ISM (tree) data etc.

  • Maintainer: Meißner Pascal <asr-ros AT lists.kit DOT edu>
  • Author: Borella Jocelyn, Hanselmann Fabian, Heller Florian, Heizmann Heinrich, Kübler Marcel, Mehlhaus Jonas, Meißner Pascal, Qattan Mohamad, Reckling Reno, Stroh Daniel
  • License: BSD
  • Source: git https://github.com/asr-ros/asr_ism.git (branch: master)

Logo-white-medium.jpg

Description

This package serves as the Passive-Scene-Recognition interface to the functionality provided by the asr_lib_ism library in the ROS environment. The Active-Scene-Recognition interface to ISM trees is located in asr_recognizer_prediction_ism, instead. The core functions of this package are the training and recognition of scenes which consist of various objects using Implicit Shape Model (ISM) trees. In addition to the implementation of the tools provided by asr_lib_ism there are several visualization (based on asr_ism_visualizations) and utility tools to make the usage of the scene recognition system more convenient.

example_recognizer_viz.png

Functionality

As mentioned above the basic functionality of the scene recognition system consists of two main parts: The training of a scene model and the recognition of a trained model in a new scene afterwards.

There are two training approaches implemented, one simply being called Trainer, the other CombinatorialTrainer. They are interchangeable but as they use different algorithms the trained models result in a difference in recognition quality mostly in terms of speed as the latter has a shorter computation time. The input of both trainers is a recording of a scene consisting of multiple objects which is stored in a sqlite database and can be created using the "Recorder" tool. The "Recognizer" tool uses the trained models and tries to find the most likely match given a set of objects.

In general the steps required are the recording of a scene, training a model from this recording and then using the recognizer to get the most likely scene match from a new object configuration. Below you can see a list of all tools this package provides including the mentioned ones and tutorials on how to use them. For more information about the functionality see the asr_lib_ism package.

Wrapper Nodes

The nodes in this list wrap tools from asr_lib_ism to provide them in the ROS environment:

  • combinatorialTrainer
  • data_cleaner
  • data_merger
  • marker_rotator
  • pose_interpolator
  • recognizer
  • recorded_objects_transformer
  • recorder
  • rotation_invariant_objects_rotator
  • trainer

Visualization Nodes

Although many nodes of this package publish some kind of visualization, the purpose of these nodes is solely to visualize ism data in rviz.

recordViewer: Visualize recorded data of a given pattern as a path for each object on which these objects moved while recording.

modelViewer: Expand the recordViewer visualization with vectors from positions of an object on its path at a certain point in time to its reference path at the same point in time. This visualization can help the user to understand how the model was created from each record data point.

voteViewer: Provide two different vote visualizations, the first just visualizes all votes from a fixed pose for a selected object in a selected subpattern and the second visualizes a given configuration with each object's votes that the model of a selected pattern provides.

Additional Nodes/Classes

ObjectConverter (class): This class can be used for custom nodes to convert ros messages (asr_msgs::AsrObject) to data-types used by the asr_lib_ism library. This class is implemented in the ism_helper.hpp/.cpp file.

scene_configurator (class): This class expands a node with the functionality to configure an object configuration interactively. Main features of this class are the multi-threaded processing of incoming messages (number of threads set by the user) and the processing of keyboard input (polling on its own dedicated thread). With this approach the actual node runs on its own thread(s) and decouples the SceneConfigurator logically and computationally from the node.

fake_data_publisher (node): This node republishes recorded data from a given database as ros messages (asr_msgs::AsrObject).

object_configuration_generator (node): This node generates an object configuration by editing object data from a given database or XML-file and writes it to an XML-file.

Usage

Because this package is a collection of nodes and helper classes to enable the user to write his own nodes, it is recommended to use the tutorials on how to use the particular tool, however in most cases the description of the nodes and their parameters should be sufficient.

Needed packages

ROS Nodes

Subscribed Topics

Published Topics

  • visualization_msgs/MarkerArray: This topic is used to visualize all kinds of data in rviz. (see visualization_topic or input_objects_visualization_topic parameters for topic name)

  • asr_msgs/AsrObject: This topic is used by the fake_data_publisher to republish object data from a given database as new object estimations. (see objectTopic parameter for topic name)

Parameters

Launch files

launch/combinatorialTrainer.launch:

  • outputDataPath (string): The path to a directory. Data such as the visualization of the ISMs created during the training process are stored here. (default: "")

  • swapRelations (bool): This parameter influences the neighbourhood of a topology T. If swapRelations == true, then topologies that can be created by swapping one relation that is part of that topology T with one that is not are considered neighbours of T. This parameter only effects TopologyGeneratorPaper. (default: true)

  • removeRelations (bool): This parameter influences the neighbourhood of a topology T. If removeRelations == true, then topologies that can be created by removing one relation from topology T are considered neighbours of T. This parameter only effects TopologyGeneratorPaper. (default: true)

  • loadValidTestSetsFrom (string): The path to the sql database that contains valid testsets. If loadValidTestSetsFrom == "" new valid testsets are generated. (default: "")

  • loadInvalidTestSetsFrom (string): The path to the sql database that contains invalid testsets. If loadInvalidTestSetsFrom == "" new invalid testsets are generated. (default: "")

  • loadStartTopologiesFrom (string): The path(s) to the file(s) that contain the topologies from which to optimization process is started. Paths must be separated by a semicolon. If loadStartTopologiesFrom is not specified for a pattern, new start topologies are generated. (default: "")

  • startFromRandomTopology (bool): If no start topology was provided, the optimization will be started from a random topology. If startFromRandomTopology == true, else the best star topology will be selected and the optimization will start from that topology. (default: true)

  • storeValidTestSetsTo (string): The path to where the valid testsets will be stored. If left empty, the testsets are not stored. (default: "")

  • storeInvalidTestSets (string): The path to where the invalid testsets will be stored. If left empty, the testsets are not stored. (default: "")

  • storeFullyMeshedISM (bool): Whether to write the ism of the fully meshed topology to a sql database or not. (default: false)

  • storeStartTopologyISM (bool): Whether to write the ism of the selected start topology to a sql database or not. (default: true)

  • testSetCount (uint): The amount of valid and invalid test sets that are generated. (default: 600)

  • confidenceThreshold (double): Threshold for how big the confidence for a recognition result must be to be accepted. (default: 0.99)

  • evaluatorId (uint): Which evaluator is used. (default: 0) (evaluatorId == 0 => Tester)

  • testForFalseNegatives (bool): Whether the test for false negatives should be run or not. False negatives have no influence on the rating of topologies. (default: false)

  • costFunctionId (uint): Which cost function is used. (default: 0) (costFunctionId == 0 => WeightedSum)

  • alpha (double): The weight for the normalised false positives. (default: 5)

  • beta (double): The weight for the normalised average recognition runtime. (default: 1)

  • topologyGeneratorId (uint): Which topologyGenerator is used (default: 0) (0 => TopologyGeneratorPaper; 1 => TopologyGeneratorNaive)

  • treeValidatorId (uint): Which treeValidator is used. (default: 0) (treeValidatorId == 0 => HeightChecker)

  • optimizationAlgorithmId (uint): Which optimization algorithm is used. (default: 2) (0 => HillClimbing; 1 => SimulatedAnnealing; 2 => RecordHunt)

  • hillClimbingStrictMonoton (bool): Whether hillclimbing continues if a new evaluation result is exactly equal to the best evaluation result (hillClimbingStrictMonoton = false) or whether a true improvement is requiered and the optimization stops if both are equal (hillClimbingStrictMonoton = true). (default: true)

  • maxTreeHeight (uint): The maximum tree height that is accepted by HeightChecker. (default: 3)

  • maxNeighbourCount (int): The maximum number of neighbours that are visited during ONE optimization step. If maxNeighbourCount == -1, all neighbours will be visited. (default: 30)

  • randomRestartProbability (double): The probability of performing a random restart. Value must be between 0 and 1. Only used if Hill Climbing is used for the optimization. (default: 0.0)

  • randomWalkProbability (double): The probability of performing a walk. Value must be between 0 and 1. Only used if Hill Climbing is used for the optimization. (default: 0.0)

  • startTemperature (double): The starting temperature of simulated annealing. Only used if Simulated Annealing is used for the optimization. (default: 1)

  • endTemperature (double): The end temperature of simulated annealing, if temperature < endTemperature simulated annealing stops. Only used if Simulated Annealing is used for the optimization. (default: 0.005)

  • temperatureFactor (double): Used to calculate the next temperature, temperature_t+1 = temperature_t * temperatureFactor. Only used if Simulated Annealing is used for the optimization. (default: 0.85)

  • repetitionsBeforeUpdated (uint): How many topologies are visited in simulated annealing before the temperature is updated. Only used if Simulated Annealing is used for the optimization. (default: 8)

  • initialAcceptableCostDelta (double): Used to initialize currentAcceptableCostDelta in RecordHunt. If new_cost is the cost we want to check and lowest_cost is the lowest cost we have encountered so far, then new_cost is accepted if: new_cost <= (1 + currentAcceptableCostDelta) * lowest_cost. Only used if Record Hunt is used for the optimzation. (default: 0.02)

  • costDeltaDecreaseFactor (double): Used to update the current costDifferenceDelta. Only used if Record Hunt is used for the optimzation. (default: 0.01)

launch/data_cleaner.launch:

  • databases (string[]): A list of databases from which the data should be cleaned/deleted. (default: empty-list)

  • clean_records (bool): Whether recordings should be deleted. (default: false).

  • clean_models (bool): Whether trained models should be deleted. (default: false).

launch/data_merger.launch:

  • target (string): Resulting merged database. (default: "")

  • sources (string[]): A list of databases which should be merged into target. (default: empty-list)

  • merge_records (bool): Whether recordings should be merged into target. (default: true)

  • merge_models (bool): Whether trained models should be merged into target. (default: true).

launch/fake_data_publisher.launch:

  • usedPattern (int): Used to pick a specific pattern from a database. (optional)

  • firstUsedSet (int): Determine the lower boundary for the used subset of recorded object datasets. (optional)

  • lastUsedSet (int): Determine the upper boundary for the used subset of recorded object datasets. (optional)

launch/marker_rotator.launch:

  • source (string): Source database. (default: "")

  • target (string): Resulting database. (default: "")

launch/modelViewer.launch:

  • sceneName (string): Name of the scene which should be visualized. (default: "")

launch/object_configuration_generator.launch:

  • object_configuration_pattern_names (string[]): List of selected patterns from database. (default: empty-list) (optional)

  • output_file_path (string): Output path for the resulting XML-file. (default: "")

  • config_file_path (string): Path to a previously created configuration XML-file. If this parameter is defined the data from the given database will be ignored. (default: "") (optional)

launch/pose_interpolator.launch:

  • source (string): Source database. (default: "")

  • target (string): Resulting interpolated database. (default: "")

  • step_number (int): Number of interpolated poses between two object poses (default: "")

launch/recognizer.launch:

  • ignoreIds (bool): Object id information in recognition result is ignored. (defaut: false)

  • ignoreTypes (bool): Object type information in recognition result is ignored. (default: false)

  • maxNumberOfResultsPerPattern (int): The maximum number of results per pattern which you want to receive from the recognizer. (default: 1)

  • raterType (int): Which rater to rate/score an recognition result should be used. (default: 0) (0 = SimpleRater: conisders only if certain conditions are met and rate it with 1 if conditions are met else with 0.; 1 = APORater: consider appearance, position and orientation for the rating. The score/rate is in [0,1].)

  • enableStoringConfigToXml (bool): Whether to store input object configurations as xml files on disk or not. Disabled if SceneConfigurator is Auto Processing. (default: false)

  • configurationFolderPath (string): Path to the folder where the configuration xml files should be stored. If left empty, no configuration files are created. (default: "")

launch/recorded_objects_transformer.launch:

  • source (string): Source database. (default: "")

  • target (string): Resulting transformed database. (default: "")

  • object_type (string): Type of a reference object. (default: "")

  • object_id (string): Id of a reference object. (default: "")

  • position (double[]): Define a reference position. (default: empty-list)

  • orientation (double[]): Define a reference orientation. (default: empty-list)

launch/recorder.launch:

  • sceneName (string): Name of the scene to which demonstration being recorded belongs. (default: test)

launch/recordViewer.launch:

  • sceneName (string): Name of the scene which should be visualized. (default: "")

launch/rotation_invariant_objects_rotator.launch:

  • source (string): Source database. (default: "")

  • target (string): Resulting transformed database. (default: "")

  • object_type (string): Type of a reference object. (default: "")

  • object_id (string): Id of a reference object. (default: "")

  • rotation_invariant_types (string[]): List of rotation invariant object-types, which should be normalized. (default: empty-list)

launch/trainer.launch:

  • useClustering (bool): Whether normal ISMs (=false) or hierarchical ISMs (=true) should be learned. (default: true)

  • dropOldModelTables (bool): Whether exiting model data should be dropped or kept. (default: true)

  • useUserDefCluster (bool): Whether user defined cluster from clusterListFile should be used. (default: false)

  • clusterListFile (string): Path to a file with user defined cluster. (default: "")

  • preDefRefListFile (string): Path to a file with predefined references for (sub-)patterns. (default: "")

  • staticBreakRatio (double): Threshold for directional heuristic. (default: 0.01)

  • togetherRatio (double): Threshold for directional heuristic. (default: 0.90)

  • maxAngleDeviation (double): Threshold for directional heuristic. (default: 45)

launch/voteViewer.launch:

  • sceneName (string): Name of the scene which should be visualized. (default: "")

  • config_file_path (string): The path to a previously created configuration XML-file. (default: "") (optional)

YAML files

param/capturing.yaml:

  • objectTopic (string): The topic a SceneConfigurator node is listening to for new object estimations. (default: /stereo/objects)

  • capture_interval (double): The interval in which a object configuration is processed when an SceneConfigurator node is set to “Auto Processing“. (default: 3)

  • queueSizePbdObjectCallback (int): The queue size for the objectTopic-Subscriber to buffer messages if the processing of new messages is slower as the time for new estimations to arrive. (default: 100)'

  • object_input_thread_count (int): The amount of threads used to process incoming object estimation messages. (default: 4)

  • use_confidence_from_msg (bool): Whether to use the confidence from the object messages or just assume perfect object recognition and use confidence 1.0 for every recognized object. (default: false)

  • enableRotationMode (int): Whether the object orientation should be rotated to the baseFrame or to an object. This only affects rotation invariant objects. (0: don't rotate; 1: rotate to rotationFrame; 2: rotate to rotationObject) (default: 1)

  • rotationFrame (string): Defines the frame to which the objects should be rotated. The rotation only takes place if baseFrame == rotationFrame and enableRotationMode = 1 . This only affects rotation invariant objects. (default: /map)

  • rotationObjectType (string): Defines the object type to which the objects should be rotated. The rotation only takes place if enableRotationMode = 2. This only affects rotation invariant objects.(default: "")

  • rotationObjectId (string): Defines the object id to which the objects should be rotated. The rotation only takes place if enableRotationMode = 2. This only affects rotation invariant objects. (default: "")

  • keyboard_poll_rate (int): The rate (Hz) in which the keyboard-thread polls the keyboard for user input. (default: 10)

  • input_objects_visualization_topic (string): The topic on which the visualization of a SceneConfigurator node is published. (default: /input_objects_scene_configurator_viz)

  • enableNeighborhoodEvaluation (bool): Whether to use neighborhood evaluation or not. (default: false)

  • angleNeighbourThreshold (double): Maximal angle deviation accepted between two estimations of the same object. Only used if neighborhood evaluation is enabled. (default: 30)

  • distanceNeighbourThreshold (double): Maximal distance accepted between two estimations of the same object. Only used if neighborhood evaluation** is enabled. (default: 0.05)

param/sensitivity.yaml:

  • bin_size (double): Side length of cubes (bins) making up a voxel grid in which hough voting is performed. Maximal accepted distance between scene reference hypotheses of different objects in a bin. (default: 0.1)

  • maxProjectionAngleDeviation (double): Maximal accepted difference in orientations of scene reference hypotheses of different objects in a bin. (default: 30)

param/sqlitedb.yaml:

  • dbfilename (string): The path to the database which should be used in the system. The recorder node will create this database if no such database exists. (default: "")

  • baseFrame (string): Frame to which incoming object messages are transformed. Used to unify their coordinate frames. (default: /map)

param/visualization.yaml:

  • visualization_topic (string): The name of the topic where the visualization markers are published to. (default: /ism_results_visualization)

Needed services

/asr_object_database/object_meta_data: This service is only called by the ObjectConverter class and provides it with additional information about recognizable objects from the asr_object_database.

Tutorials

How to use a SceneConfigurator-Node

How to use the recorder

Training an ISM

Training an ISM using the Combinatorial-Trainer

How to use the recognizer

How to use the visualization nodes

How to use the fake_data_publisher

How to use the object_configuration_generator

Troubleshooting

Problem: Missing object estimations when using a SceneConfigurator -Node and the asr_flir_ptu_driver GUI.

If you are using the ptu_gui.launch from asr_flir_ptu_driver and you checked the option “update current angle immediately”, it may happen that the SceneConfigurator clears the object estimations in the current view, because the ptu toggles its “reached_desired_position” parameter.

Solution: Make sure to uncheck the “update current angle immediately” option if you want to work with the current view.

When using ISMs, please cite the following publication. Thank you!

  • P. Meißner, R. Reckling, R. Jäkel, S. R. Schmidt-Rohr and R. Dillmann. Recognizing scenes with hierarchical implicit shape models based on spatial object relations for programming by demonstration. In International Conference on Advanced Robotics, 2013.

When using Combinatorial Trainer, please cite the following publication. Thank you!

  • P. Meißner, F. Hanselmann, R. Jäkel, S. R. Schmidt-Rohr and R. Dillmann. Automated selection of spatial object relations for modeling and recognizing indoor scenes with hierarchical implicit shape models. In International Conference on Intelligent Robots and Systems, 2015.

Wiki: asr_ism (last edited 2020-01-06 13:00:20 by PascalMeißner)