TensorParameters Struct Reference

Structure to define the input options for deep learning inference. More...

Types
enum class	COLOR_FORMAT
	Lists available color formats for the Tensor. More...
enum class	LAYOUT
	Lists available memory layouts for the Tensor. More...
enum class	PIXEL_TYPE : int
	Lists available pixel data types for the Tensor elements. More...

Functions
size_t	getPixelSize () const
	Returns the size in bytes of a single pixel element based on pixel_type. More...
int	getRequestedChannels () const
	Returns the number of channels associated with the color_format. More...

Attributes
sl::Resolution	target_size = sl::Resolution(640, 640)
	Target size for the input image (Width, Height). More...
PIXEL_TYPE	pixel_type = PIXEL_TYPE::FLOAT
	Pixel data type for the Tensor elements. More...
sl::float3	mean = {0.485f, 0.456f, 0.406f}
	Normalization parameter: Mean. More...
sl::float3	std = {0.229f, 0.224f, 0.225f}
	Normalization parameter: Standard Deviation. More...
bool	stretch = false
	If true, stretches the image to target_size. If false, keeps aspect ratio and applies letterboxing (padding). More...
sl::float3	scale = {1.f / 255.f, 1.f / 255.f, 1.f / 255.f}
	Scale factor applied to the pixel values before normalization. More...
COLOR_FORMAT	color_format = COLOR_FORMAT::BGR
	Desired color format for the output Tensor. More...
LAYOUT	layout = LAYOUT::NCHW
	Desired memory layout for the output Tensor. More...
size_t	batch_size = 1
	Number of images in the batch. More...
MEM	memory_type = MEM::GPU
	Where the Tensor should be stored (CPU or GPU). More...

Detailed Description

Structure to define the input options for deep learning inference.

Enumeration

COLOR_FORMAT

enum COLOR_FORMAT

strong

Lists available color formats for the Tensor.

Enumerator
GRAY	1-channel Grayscale.
Y	1-channel Y (Luminance) from YUV.
RGB	3-channel RGB.
BGR	3-channel BGR.
RGBA	4-channel RGBA.
BGRA	4-channel BGRA.

LAYOUT

enum LAYOUT

strong

Lists available memory layouts for the Tensor.

Enumerator
NCHW	Batch, Channels, Height, Width (Planar). Common for PyTorch models.
NHWC	Batch, Height, Width, Channels (Interleaved). Common for TensorFlow models.

PIXEL_TYPE

enum PIXEL_TYPE : int

strong

Lists available pixel data types for the Tensor elements.

Enumerator
FLOAT	32-bit floating point.
HALF	16-bit floating point (half precision).
UCHAR	Unsigned char (8-bit).

Functions

getPixelSize()

size_t getPixelSize ( ) const

inline

Returns the size in bytes of a single pixel element based on pixel_type.

Returns

The size in bytes (1 for UCHAR, 2 for HALF, 4 for FLOAT).

getRequestedChannels()

int getRequestedChannels ( ) const

inline

Returns the number of channels associated with the color_format.

Returns

The number of channels.

Variables

target_size

sl::Resolution target_size = sl::Resolution(640, 640)

Target size for the input image (Width, Height).

pixel_type

PIXEL_TYPE pixel_type = PIXEL_TYPE::FLOAT

Pixel data type for the Tensor elements.

Note

Specifies the underlying data type: UCHAR (8-bit), FLOAT (32-bit), or HALF (16-bit).

mean

sl::float3 mean = {0.485f, 0.456f, 0.406f}

Normalization parameter: Mean.

Note

Pixel = (Pixel * Scale - Mean) / Std.

Default: ImageNet Standard {0.485, 0.456, 0.406}.

std

sl::float3 std = {0.229f, 0.224f, 0.225f}

Normalization parameter: Standard Deviation.

Note

Pixel = (Pixel * Scale - Mean) / Std.

Default: ImageNet Standard {0.229, 0.224, 0.225}.

stretch

bool stretch = false

If true, stretches the image to target_size. If false, keeps aspect ratio and applies letterboxing (padding).

scale

sl::float3 scale = {1.f / 255.f, 1.f / 255.f, 1.f / 255.f}

Scale factor applied to the pixel values before normalization.

color_format

COLOR_FORMAT color_format = COLOR_FORMAT::BGR

Desired color format for the output Tensor.

layout

LAYOUT layout = LAYOUT::NCHW

Desired memory layout for the output Tensor.

Note

NCHW (planar) is common for PyTorch models.

NHWC (interleaved) is common for TensorFlow models.

batch_size

size_t batch_size = 1

Number of images in the batch.

memory_type

MEM memory_type = MEM::GPU

Where the Tensor should be stored (CPU or GPU).