AdaBoostTrigger Machine Learning
The AdaBoostTrigger is a binary technology type. It uses the Adaptive Boosting (AdaBoost) machine learning algorithm to determine when a player performs a certain gesture.
During training time it accepts input tags, boolean values, which mark the occurrence of an event, such as a hit. This marking or tagging is used to evaluate whether or not a gesture has happened and determines the confidence value of the event.
Note You can override the AdaBoostTrigger project settings in the solution global settings.
Input Parameters
The following table describes the input parameters that you can use when tagging a gesture. You enter these parameters in the Project Settings grid, as shown in the Visual Gesture Builder Training Project.
Parameters marked (New) are new to the latest release. If you have an earlier VGB project, the project will be updated with the latest properties the first time when it is opened in the latest release, and you will be prompted to save the changes to the updated project.
| Property Name | Type | Description | ||
|---|---|---|---|---|
| Accuracy Level | FLOAT | A floating point value in the range [0..1]. This value controls how accurate your results are, but also affects the training time. The higher the accuracy, the longer the training time. As a guideline you can use the following values:
| ||
| Number of Weak Classifiers at Runtime | INT | The algorithm can potentially generate tens of thousands of weak classifiers. Using all of them will increase accuracy, but at the higher CPU cost on the console. To use all the classifiers generated, use a value of 0. The CPU cost for 1000 weak classifiers is only 25 microseconds and the results have more than adequate accuracy. | ||
| Filter Results | BOOL | The results of the algorithm are based on a frame, not on a gesture. A filter needs to be applied on the raw per frame results. The AdaBoostTrigger provides a simple low latency filter, but you have the option of disabling filtering and applying your own filter at runtime on the console. The filter used is a simple sliding window of N frames, summing up the results and comparing it against a threshold value. The number of frames can be seen as a frequency and the threshold can be seen as amplitude. | ||
| Auto Find Best Filtering Params | BOOL | When filtering is switched on, you have the option to let the trainer automatically find best filtering parameters which will minimize the rate of false positives and false negatives | ||
| Weight Of False Positives During Auto Find | FLOAT | A value in the range of [0..1] which is used when automatically finding the best filtering parameters.If it is more important to reduce false positives, use a higher value. If it is more important to reduce false negatives, use a lower value. | ||
| Manual Filter Params: Number of Frames To Filter | INT | A value in the range of [1..10].You can choose the number of frames over which to filter if you do not choose to find the best filtering parameters automatically. | ||
| Manual Filter Params: Threshold | FLOAT | A value in the range of [0..1].Lower values could increase true positives, at the risk of increasing false positives.Higher values could decrease false positives, at the risk of decreasing true positives. | ||
| Duplicate And Mirror Data During Training | BOOL | Skeleton data can be duplicated and mirrored in order to have a larger set of training data. | ||
| % CPU For Training | INT | A value in the range of [0..100] which indicates the % of CPU resources the trainer should use for training. | ||
| Use Hands Data (New) | BOOL | Default to false. This means by default the hand states are not used for training and detection. For training and detection to use the hand states, set this property to true.
| ||
| Ignore Left Arm (New) | BOOL | Default to false. This means by default the following left-arm joints are used for training.
| ||
| Ignore Right Arm (New) | BOOL | Default to false. This means by default the following right-arm joints are used for training.
| ||
| Ignore Lower Body (New) | BOOL | Default to false. This means by default the following lower-body joints are used for training.
|
Note Runtime Data
An AdaBoostTrigger gesture will be represented as a GestureType.Discrete Gesture during runtime. The DiscreteGestureResult Class should be used to check if the gesture has been detected and at what confidence.
Remarks
In training, if you only tag positive frames, AdaBoostTrigger uses all frames by default. The frames not tagged are used as negative frames. This generates a big training set, if the clips are very long.
In training, if you intend to use only some frame ranges, instead of all the frames of a clip, you can explicitly tag negative frames. During training, the tool detects these explicitly tagged negative frames and ignores the non-tagged frames.
The tool UI is designed to help you convert from implicit negative to explicit negative tagging.
When you select a range of frames and set its value to FALSE, if there are non-tagged frames (gap) in the range, only those gaps are set to FALSE. All existing tagged frames are not affected.
However, the range works like a bundle set to FALSE in its entirety, if there is no gap.