Gaussian Image Filter

Overview

Gaussian image filter is a low-pass discrete Gaussian filter that smooths out the image by doing a Gaussian-weighted averaging of neighbor pixels of a given input pixel. It produces images with less artifacts than Box Image Filter, but could potentially be more costly to compute.

It supports two modes of operation:

• Kernel support size is automatically calculated based on the filter standard deviation (sigma).
• Use both user-provided kernel support size and filter standard deviation.

Input	Gaussian kernel	Output
	7x7 support, \[ \sigma=1.7 \]

Implementation

Gaussian filter is implemented as a convolution operation on the input image where the kernel has the following weights:

\[ w_g[x,y] = \frac{1}{2\pi\sigma^2} \cdot e^{-\frac{x^2+y^2}{2\sigma^2}} \]

When the input kernel support size is 0 for a given dimension (or both), it is calculated from the given standard deviation by assuming that the weights outside \(\pm3\sigma\) window are zero.

In this case, the following formula is used:

\[ w = \max\{3,2 \times \lceil 3\sigma\rceil-1\} \]

Note

We clamp the minimum kernel size to 3 because a kernel with size 1 doesn't have enough samples to properly characterize a Gaussian function.

Usage

1. Initialization phase
   1. Include the header that defines the Gaussian filter function.

      #include <vpi/algo/GaussianImageFilter.h>

   2. Define the stream on which the algorithm will be executed, the input and output images.

          VPIStream stream = /*...*/;
          VPIImage input   = /*...*/;

   3. Create the output image.

          uint32_t w, h;
          vpiImageGetSize(input, &w, &h);
       
          VPIImageType type;
          vpiImageGetType(input, &type);
       
          VPIImage output;
          vpiImageCreate(w, h, type, 0, &output);

2. Processing phase
   1. Submit the algorithm to the stream, input, output images, window size and boundary condition.

          vpiSubmitGaussianImageFilter(stream, input, output, 7, 7, 1.7, 1.7, VPI_BOUNDARY_COND_ZERO);

   2. Optionally, wait until the processing is done.

          vpiStreamSync(stream);

Limitations and Constraints

Constraints for specific backends supersede the ones specified for all backends.

All Backends

• Kernel size must be at least 1x1 and at most 11x11.
• Accepted boundary conditions:
  • VPI_BOUNDARY_COND_CLAMP
  • VPI_BOUNDARY_COND_ZERO
• Input and output images must have same image type and size.
• Accepted input image type:
  • VPI_IMAGE_TYPE_U8
  • VPI_IMAGE_TYPE_S8
  • VPI_IMAGE_TYPE_U16
  • VPI_IMAGE_TYPE_S16

PVA

• Minimum kernel size is 2x2.
• Accepted boundary conditions:
  • VPI_BOUNDARY_COND_ZERO

Performance

For further information on how performance was benchmarked, see Performance Measurement.

Jetson AGX Xavier

size	type	kernel	CPU	CUDA	PVA
1920x1080	u8	3x3	0.48 ms	0.0649 ms	0.9617 ms
1920x1080	u8	5x5	0.611 ms	0.0689 ms	1.2313 ms
1920x1080	u8	7x7	0.63 ms	0.0881 ms	1.7580 ms
1920x1080	u8	11x11	0.860 ms	0.0995 ms	3.2604 ms
1920x1080	u16	3x3	0.43 ms	0.1077 ms	1.0608 ms
1920x1080	u16	5x5	0.59 ms	0.1163 ms	1.5124 ms
1920x1080	u16	7x7	1.12 ms	0.1340 ms	2.3514 ms
1920x1080	u16	11x11	1.31 ms	0.1586 ms	4.6484 ms

Jetson TX2

size	type	kernel	CPU	CUDA	PVA
1920x1080	u8	3x3	0.89 ms	0.254 ms	n/a
1920x1080	u8	5x5	1.013 ms	0.288 ms	n/a
1920x1080	u8	7x7	1.21 ms	0.396 ms	n/a
1920x1080	u8	11x11	1.57 ms	0.471 ms	n/a
1920x1080	u16	3x3	1.81 ms	0.375 ms	n/a
1920x1080	u16	5x5	2.15 ms	0.418 ms	n/a
1920x1080	u16	7x7	2.57 ms	0.577 ms	n/a
1920x1080	u16	11x11	3.41 ms	0.668 ms	n/a

Jetson Nano

size	type	kernel	CPU	CUDA	PVA
1920x1080	u8	3x3	1.523 ms	0.6650 ms	n/a
1920x1080	u8	5x5	2.253 ms	0.742 ms	n/a
1920x1080	u8	7x7	2.71 ms	1.024 ms	n/a
1920x1080	u8	11x11	3.004 ms	1.235 ms	n/a
1920x1080	u16	3x3	3.662 ms	0.965 ms	n/a
1920x1080	u16	5x5	4.227 ms	1.015 ms	n/a
1920x1080	u16	7x7	5.29 ms	1.389 ms	n/a
1920x1080	u16	11x11	7.435 ms	1.657 ms	n/a