Detailed Description

Defines ObjectDetector module for object detection.

Data Structures
struct	dwObjectDetectorDNNParams

struct	dwObjectDetectorParams

Macros
#define	DW_OBJECT_DETECTOR_MAX_IMAGES 16
	Maximum number of images to run detector on. More...

Typedefs
typedef struct dwObjectDetectorObject *	dwObjectDetectorHandle_t
	Handle to an object detector. More...

Functions
DW_API_PUBLIC dwStatus	dwObjectDetector_getCoverageThreshold (float32_t *covThreshold, uint32_t classIdx, dwObjectDetectorHandle_t obj)
	Get coverage threshold for a given class. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_getCUDAStream (cudaStream_t *stream, dwObjectDetectorHandle_t obj)
	Gets the CUDA stream used. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_getDetectedObjects (dwObject objectList, size_t numObjects, uint32_t imageIdx, uint32_t classIdx, dwObjectDetectorHandle_t obj)
	Returns detected objects for one of the images given to dwObjectDetector_inferDeviceAsync. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_getROI (dwRect ROI, dwTransformation2D outputToObject, uint32_t imageIdx, dwObjectDetectorHandle_t obj)
	Returns the region of interest and the 2D transformation from the image coordinate system to object coordinate system at given index. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_inferDeviceAsync (const dwImageCUDA const imageArray, uint32_t numImages, dwObjectDetectorHandle_t obj)
	Runs inference using the provided DNN model. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_initDefaultParams (dwObjectDetectorDNNParams dnnParams, dwObjectDetectorParams detectorParams)
	Initializes ObjectDetector parameters with default values. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_initialize (dwObjectDetectorHandle_t obj, dwContextHandle_t ctx, dwDNNHandle_t dnn, const dwObjectDetectorDNNParams dnnParams, const dwObjectDetectorParams *detectorParams)
	Initializes ObjectDetector module with a preloaded DNN inference module. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_initializeFromDriveNet (dwObjectDetectorHandle_t obj, dwContextHandle_t ctx, dwDriveNetHandle_t drivenet, const dwObjectDetectorParams detectorParams)
	Initializes ObjectDetector module with a DriveNet module. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_interpretHost (uint32_t numImages, dwObjectDetectorHandle_t obj)
	Output of the inference (dwObjectDetector_inferDeviceAsync) is interpreted to get the list of object proposals. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_isFuseObjectsEnabled (dwBool *enabled, dwObjectDetectorHandle_t obj)
	Returns a boolean indicating whether object fusing is enabled. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_release (dwObjectDetectorHandle_t *obj)
	Releases the ObjectDetector module. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_reset (dwObjectDetectorHandle_t obj)
	Resets ObjectDetector. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_setCoverageThreshold (float32_t covThreshold, uint32_t classIdx, dwObjectDetectorHandle_t obj)
	Set coverage threshold for a given class. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_setCUDAStream (cudaStream_t stream, dwObjectDetectorHandle_t obj)
	Sets the CUDA stream used. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_setEnableFuseObjects (dwBool enable, dwObjectDetectorHandle_t obj)
	Enables or disables objectFusing. More...

DW_API_PUBLIC dwStatus	dwObjectDetector_setROI (uint32_t imageIdx, const dwRect ROI, const dwTransformation2D outputToObject, dwObjectDetectorHandle_t obj)
	Sets the region of interest for an image at given index, and the 2D transformation from the image coordinate system to object coordinate system. More...

Data Structure Documentation

◆ dwObjectDetectorDNNParams

struct dwObjectDetectorDNNParams

Data Fields
const char *	bottomVisBlobName	Name of the bottom visibility layer. Default is empty string.
const char *	boundingBoxBlobName	Name of the bounding box layer. Default is empty string.
float32_t	boxScaleX1	Scale coefficient for X coordinate of upper-left corner of the RAW bounding box from DNN output. Default is 1.0.
float32_t	boxScaleX2	Scale coefficient for X coordinate of lower-right corner of the RAW bounding box from DNN output. Default is 1.0.
float32_t	boxScaleY1	Scale coefficient for Y coordinates of upper-left corner of the RAW bounding box from DNN output. Default is 1.0
float32_t	boxScaleY2	Scale coefficient for Y coordinates of lower-right corner of the RAW bounding box from DNN output. Default is 1.0.
const char *	coverageBlobName	Name of the coverage layer. Default is empty string.
float32_t	coverageThreshold[DW_OBJECT_MAX_CLASSES]	Threshold for coverage for each class. Default is 0.5f for all classes.
dwDataConditionerParams	dataConditionerParams	DataConditioner parameters.
dwBool	enable2_5D	Whether the network can predict object orientation.
const char *	horizontalVisBlobName	Name of the horizontal visibility layer. Default is empty string.
uint32_t	maxProposalsPerClass	Maximum number of RAW detections directly obtained from DNN. Default is 100.
const char *	orientationYawBlobName	Name of the orientation yaw layer. Default is empty string.

◆ dwObjectDetectorParams

struct dwObjectDetectorParams

Data Fields
dwBool	enableBoundingBoxClipping	If enableBoundingBoxClipping is set to DW_TRUE, the bounding boxes are clipped to the image edges. Default is DW_TRUE since it has been used during training.
dwBool	enableFuseObjects	If enableFuseObjects is set to DW_TRUE, all the objects proposals coming from all the images are treated as a single list of proposals. Clustering is applied to this entire list, which results in a single list of objects. Default is DW_FALSE.
uint32_t	maxNumImages	Maximum number of images (cannot be larger than DW_DETECTOR_MAX_IMAGES). Default is 1.
dwRect	ROIs[DW_OBJECT_DETECTOR_MAX_IMAGES]	Region of interest for each image at the corresponding index. The image is cropped at the given ROI. The cropped region is then be scaled while keeping the aspect ratio and is padded to have the right size for DNN. Default ROI is the whole image.
dwTransformation2D	transformations[DW_OBJECT_DETECTOR_MAX_IMAGES]	List of transformations from each image coordinate system to object coordinate system at the corresponding image index. By default, the transformation matrices are identity.

Macro Definition Documentation

◆ DW_OBJECT_DETECTOR_MAX_IMAGES

#define DW_OBJECT_DETECTOR_MAX_IMAGES 16

Maximum number of images to run detector on.

Definition at line 61 of file Detector.h.

Typedef Documentation

◆ dwObjectDetectorHandle_t

typedef struct dwObjectDetectorObject* dwObjectDetectorHandle_t

Handle to an object detector.

Definition at line 67 of file Detector.h.

Function Documentation

◆ dwObjectDetector_getCoverageThreshold()

DW_API_PUBLIC dwStatus dwObjectDetector_getCoverageThreshold	(	float32_t *	covThreshold,
		uint32_t	classIdx,
		dwObjectDetectorHandle_t	obj
	)

Get coverage threshold for a given class.

Parameters

[out]	covThreshold	Coverage threshold in [0.0, 1.0]
[in]	classIdx	Index of class for which the coverage threshold should be returned.
[in]	obj	A pointer to the ObjectDetector handle.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST, DW_INVALID_ARGUMENT

◆ dwObjectDetector_getCUDAStream()

DW_API_PUBLIC dwStatus dwObjectDetector_getCUDAStream	(	cudaStream_t *	stream,
		dwObjectDetectorHandle_t	obj
	)

Gets the CUDA stream used.

Parameters

[out]	stream	The CUDA stream used.
[in]	obj	A pointer to the ObjectDetector handle.

Returns: DW_INVALID_ARGUMENT, DW_SUCCESS

◆ dwObjectDetector_getDetectedObjects()

DW_API_PUBLIC dwStatus dwObjectDetector_getDetectedObjects	(	dwObject *	objectList,
		size_t *	numObjects,
		uint32_t	imageIdx,
		uint32_t	classIdx,
		dwObjectDetectorHandle_t	obj
	)

Returns detected objects for one of the images given to dwObjectDetector_inferDeviceAsync.

Parameters

[out]	objectList	Object list where the detected objects are copied into.
[out]	numObjects	Number of objects that are detected for this image and class.
[in]	imageIdx	Index of the image the array as given to dwObjectDetector_inferDeviceAsync.
[in]	classIdx	Class index.
[in]	obj	Specifies the ObjectDetector handle.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST, DW_INVALID_ARGUMENT

Note: objectList must be allocated with maxProposalsPerClass elements.

◆ dwObjectDetector_getROI()

DW_API_PUBLIC dwStatus dwObjectDetector_getROI	(	dwRect *	ROI,
		dwTransformation2D *	outputToObject,
		uint32_t	imageIdx,
		dwObjectDetectorHandle_t	obj
	)

Returns the region of interest and the 2D transformation from the image coordinate system to object coordinate system at given index.

Parameters

[out]	ROI	Region of interest to be fed as input to network.
[out]	outputToObject	2D Transformation from output of the network to object coordinate system.
[in]	imageIdx	Index of the image for which to set ROI and transformation.
[in]	obj	Specifies the ObjectDetector handle.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST, DW_INVALID_ARGUMENT

Note: If multiple regions for a single image are to given, the image array that's given as an input to dwObjectDetector_inferDeviceAsync must include the same image at the corresponding indices.

◆ dwObjectDetector_inferDeviceAsync()

DW_API_PUBLIC dwStatus dwObjectDetector_inferDeviceAsync	(	const dwImageCUDA const	imageArray,
		uint32_t	numImages,
		dwObjectDetectorHandle_t	obj
	)

Runs inference using the provided DNN model.

The inference pipeline is as follows:

DataPreparation: Input images are prepared by scaling, padding and other transformations set in dataConditionerParams at initialization time in order to have the right properties required by the DNN. If the given DNN model expects batched input, the given images are divided into batches, where each batch has the expected batch size.
Inference: Inference is run on the prepared images / batches.

Parameters

[in]	imageArray	Array storing pointers to images where the detector is to be applied.
[in]	numImages	Number of images provided in the array.
[in]	obj	Specifies the ObjectDetector handle.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST, DW_INVALID_ARGUMENT

Note: numImages must be divisible by the batch size of the network.

◆ dwObjectDetector_initDefaultParams()

DW_API_PUBLIC dwStatus dwObjectDetector_initDefaultParams	(	dwObjectDetectorDNNParams *	dnnParams,
		dwObjectDetectorParams *	detectorParams
	)

Initializes ObjectDetector parameters with default values.

Parameters

[out]	dnnParams	ObjectDetector DNN parameters.
[out]	detectorParams	ObjectDetector parameters.

Returns: DW_SUCCESS, DW_INVALID_ARGUMENT

◆ dwObjectDetector_initialize()

DW_API_PUBLIC dwStatus dwObjectDetector_initialize	(	dwObjectDetectorHandle_t *	obj,
		dwContextHandle_t	ctx,
		dwDNNHandle_t	dnn,
		const dwObjectDetectorDNNParams *	dnnParams,
		const dwObjectDetectorParams *	detectorParams
	)

Initializes ObjectDetector module with a preloaded DNN inference module.

Parameters

[out]	obj	A pointer to ObjectDetector handle that is initialized from parameters.
[in]	ctx	Specifies the handle to the context.
[in]	dnn	Specifies the handle to the dnn module which is initialized with detector network.
[in]	dnnParams	A pointer to ObjectDetectorDNN parameters.
[in]	detectorParams	A pointer to ObjectDetector parameters.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST, DW_INVALID_ARGUMENT

Note: DNN model is expected to provide two output blobs: A blob for coverage and a blob for bounding box.

The number of channels in the coverage blob is equal to the number of classes that the network detects.

The number of channels in the bounding box blob is equal to the 4 times number of classes that the network detects.

For example, for detecting cars and pedestrians, the first channel of the coverage blob refers to the coverage of cars, and the first four channels of the bounding box blob refers to the top left and bottom right corners of car bounding boxes. The second channel of the coverage blob refers to the coverage of pedestrians, and the second four channels of the bounding box blob refers to the top left and bottom right corners of pedestrian bounding boxes.

Since it is unknown to the model which channel refers to which class, they are assigned class IDs instead, where class ID refers to the channel index of the network output. In the above example, the class ID of the car objects is 0, whereas the class ID of pedestrian objects is 1.

◆ dwObjectDetector_initializeFromDriveNet()

DW_API_PUBLIC dwStatus dwObjectDetector_initializeFromDriveNet	(	dwObjectDetectorHandle_t *	obj,
		dwContextHandle_t	ctx,
		dwDriveNetHandle_t	drivenet,
		const dwObjectDetectorParams *	detectorParams
	)

Initializes ObjectDetector module with a DriveNet module.

Parameters

[out]	obj	A pointer to ObjectDetector handle that is initialized from parameters.
[in]	ctx	Specifies the handle to the context.
[in]	drivenet	Specifies the handle to the DriveNet module.
[in]	detectorParams	A pointer to ObjectDetector parameters.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST, DW_INVALID_ARGUMENT

◆ dwObjectDetector_interpretHost()

DW_API_PUBLIC dwStatus dwObjectDetector_interpretHost	(	uint32_t	numImages,
		dwObjectDetectorHandle_t	obj
	)

Output of the inference (dwObjectDetector_inferDeviceAsync) is interpreted to get the list of object proposals.

If fusing is activated, proposals coming from different images is filled into a single list.

Parameters

[in]	numImages	Number of images provided in the array in dwObjectDetector_inferDeviceAsync call.
[in]	obj	Specifies the ObjectDetector handle.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST, DW_INVALID_ARGUMENT

◆ dwObjectDetector_isFuseObjectsEnabled()

DW_API_PUBLIC dwStatus dwObjectDetector_isFuseObjectsEnabled	(	dwBool *	enabled,
		dwObjectDetectorHandle_t	obj
	)

Returns a boolean indicating whether object fusing is enabled.

Parameters

[out]	enabled	Boolean indicating whether object fusing is enabled.
[in]	obj	A pointer to the ObjectDetector handle.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST, DW_INVALID_ARGUMENT

◆ dwObjectDetector_release()

DW_API_PUBLIC dwStatus dwObjectDetector_release ( dwObjectDetectorHandle_t * obj )

Releases the ObjectDetector module.

Parameters

[in] obj A pointer to the object handle to release.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST

Note: This method renders the handle unusable.

◆ dwObjectDetector_reset()

DW_API_PUBLIC dwStatus dwObjectDetector_reset ( dwObjectDetectorHandle_t obj )

Resets ObjectDetector.

Parameters

[in] obj Handle to reset.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST

◆ dwObjectDetector_setCoverageThreshold()

DW_API_PUBLIC dwStatus dwObjectDetector_setCoverageThreshold	(	float32_t	covThreshold,
		uint32_t	classIdx,
		dwObjectDetectorHandle_t	obj
	)

Set coverage threshold for a given class.

Parameters

[in]	covThreshold	Coverage threshold in [0.0, 1.0]
[in]	classIdx	Index of class for which the coverage threshold should be set.
[in]	obj	A pointer to the ObjectDetector handle that is updated

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST

◆ dwObjectDetector_setCUDAStream()

DW_API_PUBLIC dwStatus dwObjectDetector_setCUDAStream	(	cudaStream_t	stream,
		dwObjectDetectorHandle_t	obj
	)

Sets the CUDA stream used.

Parameters

[in]	stream	The CUDA stream used.
[in]	obj	A pointer to the ObjectDetector handle that is updated.

Returns: DW_INVALID_ARGUMENT, DW_SUCCESS

◆ dwObjectDetector_setEnableFuseObjects()

DW_API_PUBLIC dwStatus dwObjectDetector_setEnableFuseObjects	(	dwBool	enable,
		dwObjectDetectorHandle_t	obj
	)

Enables or disables objectFusing.

Parameters

[in]	enable	Boolean indicating whether object fusing should be enabled.
[in]	obj	A pointer to the ObjectDetector handle that is updated.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST

◆ dwObjectDetector_setROI()

DW_API_PUBLIC dwStatus dwObjectDetector_setROI	(	uint32_t	imageIdx,
		const dwRect *	ROI,
		const dwTransformation2D *	outputToObject,
		dwObjectDetectorHandle_t	obj
	)

Sets the region of interest for an image at given index, and the 2D transformation from the image coordinate system to object coordinate system.

Parameters

[in]	imageIdx	Index of the image for which to set ROI and transformation.
[in]	ROI	Region of interest to be fed as input to network.
[in]	outputToObject	2D Transformation from output of the network to object coordinate system.
[in]	obj	Specifies the ObjectDetector handle.

Returns: DW_SUCCESS, DW_INVALID_HANDLE, DW_BAD_CAST, DW_INVALID_ARGUMENT

Note: If multiple regions for a single image are to be given, the image array that's given as an input to dwObjectDetector_inferDeviceAsync must include the same image at the corresponding indices.

DriveWorks SDK Reference

| 0.6.67 Release

Detailed Description

Data Structures

Macros

Typedefs

Functions

Data Structure Documentation

◆ dwObjectDetectorDNNParams

◆ dwObjectDetectorParams

Macro Definition Documentation

◆ DW_OBJECT_DETECTOR_MAX_IMAGES

Typedef Documentation

◆ dwObjectDetectorHandle_t

Function Documentation

◆ dwObjectDetector_getCoverageThreshold()

◆ dwObjectDetector_getCUDAStream()

◆ dwObjectDetector_getDetectedObjects()

◆ dwObjectDetector_getROI()

◆ dwObjectDetector_inferDeviceAsync()

◆ dwObjectDetector_initDefaultParams()

◆ dwObjectDetector_initialize()

◆ dwObjectDetector_initializeFromDriveNet()

◆ dwObjectDetector_interpretHost()

◆ dwObjectDetector_isFuseObjectsEnabled()

◆ dwObjectDetector_release()

◆ dwObjectDetector_reset()

◆ dwObjectDetector_setCoverageThreshold()

◆ dwObjectDetector_setCUDAStream()

◆ dwObjectDetector_setEnableFuseObjects()

◆ dwObjectDetector_setROI()