Structure containing a set of parameters for the object detection module. More...

Data Fields
unsigned int	instance_module_id
	Id of the module instance. More...

bool	enable_tracking
	Whether the object detection system includes object tracking capabilities across a sequence of images. More...

bool	enable_segmentation
	Whether the object masks will be computed. More...

enum SL_OBJECT_DETECTION_MODEL	detection_model
	SL_OBJECT_DETECTION_MODEL to use. More...

char *	fused_objects_group_name
	In a multi camera setup, specify which group this model belongs to. More...

char *	custom_onnx_file
	Path to the YOLO-like onnx file for custom object detection ran in the ZED SDK. More...

struct SL_Resolution	custom_onnx_dynamic_input_shape
	Resolution to the YOLO-like onnx file for custom object detection ran in the ZED SDK. This resolution defines the input tensor size for dynamic shape ONNX model only. The batch and channel dimensions are automatically handled, it assumes it's color images like default YOLO models. More...

float	max_range
	Upper depth range for detections. More...

struct SL_BatchParameters	batch_parameters
	Batching system parameters. More...

enum SL_OBJECT_FILTERING_MODE	filtering_mode
	Filtering mode that should be applied to raw detections. More...

float	prediction_timeout_s
	Prediction duration of the ZED SDK when an object is not detected anymore before switching its state to SL_OBJECT_TRACKING_STATE_SEARCHING. More...

bool	allow_reduced_precision_inference
	Whether to allow inference to run at a lower precision to improve runtime and memory usage. More...

Detailed Description

Structure containing a set of parameters for the object detection module.

The default constructor sets all parameters to their default settings.

Note: Parameters can be user adjusted.

Field Documentation

◆ instance_module_id

unsigned int SL_ObjectDetectionParameters::instance_module_id

Id of the module instance.

This is used to identify which object detection module instance is used.

◆ enable_tracking

bool SL_ObjectDetectionParameters::enable_tracking

Whether the object detection system includes object tracking capabilities across a sequence of images.

Default: true

◆ enable_segmentation

bool SL_ObjectDetectionParameters::enable_segmentation

Whether the object masks will be computed.

Default: false

◆ detection_model

enum SL_OBJECT_DETECTION_MODEL SL_ObjectDetectionParameters::detection_model

SL_OBJECT_DETECTION_MODEL to use.

Default: SL_OBJECT_DETECTION_MODEL_MULTI_CLASS_BOX_FAST

◆ fused_objects_group_name

char* SL_ObjectDetectionParameters::fused_objects_group_name

In a multi camera setup, specify which group this model belongs to.

In a multi camera setup, multiple cameras can be used to detect objects and multiple detector having similar output layout can see the same object. Therefore, Fusion will fuse together the outputs received by multiple detectors only if they are part of the same fused_objects_group_name.

Note: This parameter is not used when not using a multi-camera setup and must be set in a multi camera setup.

◆ custom_onnx_file

char* SL_ObjectDetectionParameters::custom_onnx_file

Path to the YOLO-like onnx file for custom object detection ran in the ZED SDK.

When detection_model is OBJECT_DETECTION_MODEL::CUSTOM_YOLOLIKE_BOX_OBJECTS, a onnx model must be passed so that the ZED SDK can optimize it for your GPU and run inference on it.

The resulting optimized model will be saved for re-use in the future.

Attention: - The model must be a YOLO-like model.; - The caching uses the custom_onnx_file string along with your GPU specs to decide whether to use the cached optmized model or to optimize the passed onnx model. If you want to use a different model (i.e. an onnx with different weights), you must use a different custom_onnx_file string or delete the cached optimized model in <ZED Installation path>/resources.

Note: This parameter is useless when detection_model is not OBJECT_DETECTION_MODEL::CUSTOM_YOLOLIKE_BOX_OBJECTS.

◆ custom_onnx_dynamic_input_shape

struct SL_Resolution SL_ObjectDetectionParameters::custom_onnx_dynamic_input_shape

Resolution to the YOLO-like onnx file for custom object detection ran in the ZED SDK. This resolution defines the input tensor size for dynamic shape ONNX model only. The batch and channel dimensions are automatically handled, it assumes it's color images like default YOLO models.

Note: This parameter is only used when detection_model is OBJECT_DETECTION_MODEL::CUSTOM_YOLOLIKE_BOX_OBJECTS and the provided ONNX file is using dynamic shapes.

Attention: - Multiple model only support squared images

\default Squared images 512x512 (input tensor will be 1x3x512x512)

◆ max_range

float SL_ObjectDetectionParameters::max_range

Upper depth range for detections.

Default: -1.f (value set in SL_InitParameters.depth_maximum_distance)

Note: The value cannot be greater than SL_InitParameters.depth_maximum_distance and its unit is defined in SL_InitParameters.coordinate_unit.

◆ batch_parameters

struct SL_BatchParameters SL_ObjectDetectionParameters::batch_parameters

Batching system parameters.

Batching system (introduced in 3.5) performs short-term re-identification with deep-learning and trajectories filtering.
SL_BatchParameters.enable must to be true to use this feature (by default disabled).

◆ filtering_mode

enum SL_OBJECT_FILTERING_MODE SL_ObjectDetectionParameters::filtering_mode

Filtering mode that should be applied to raw detections.

Default: SL_OBJECT_FILTERING_MODE_NMS_3D (same behavior as previous ZED SDK version)

Note: This parameter is only used in detection model SL_OBJECT_DETECTION_MODEL_MULTI_CLASS_BOX_XXX and SL_OBJECT_DETECTION_MODEL_CUSTOM_BOX_OBJECTS.; For custom object, it is recommended to use SL_OBJECT_FILTERING_MODE_NMS_3D_PER_CLASS or SL_OBJECT_FILTERING_MODE_NONE.; In this case, you might need to add your own NMS filter before ingesting the boxes into the object detection module.

◆ prediction_timeout_s

float SL_ObjectDetectionParameters::prediction_timeout_s

Prediction duration of the ZED SDK when an object is not detected anymore before switching its state to SL_OBJECT_TRACKING_STATE_SEARCHING.

It prevents the jittering of the object state when there is a short misdetection.
The user can define their own prediction time duration.
Default: 0.2f

Note: During this time, the object will have SL_OBJECT_TRACKING_STATE_OK state even if it is not detected.; The duration is expressed in seconds.

Warning: prediction_timeout_s will be clamped to 1 second as the prediction is getting worse with time.; Setting this parameter to 0 disables the ZED SDK predictions.

◆ allow_reduced_precision_inference

bool SL_ObjectDetectionParameters::allow_reduced_precision_inference

Whether to allow inference to run at a lower precision to improve runtime and memory usage.

It might increase the initial optimization time and could include downloading calibration data or calibration cache and slightly reduce the accuracy.

Note: The fp16 is automatically enabled if the GPU is compatible and provides a speed up of almost x2 and reduce memory usage by almost half, no precision loss.; This setting allow int8 precision which can speed up by another x2 factor (compared to fp16, or x4 compared to fp32) and half the fp16 memory usage, however some accuracy could be lost.; The accuracy loss should not exceed 1-2% on the compatible models.; The current compatible models are all SL_AI_MODELS_HUMAN_BODY_XXXX.

Data Fields