Merge pull request #15942 from OrestChura:fb_tutorial

[platform/upstream/opencv.git] / doc / tutorials / gapi / face_beautification / face_beautification.markdown

# Implementing a face beautification algorithm with G-API {#tutorial_gapi_face_beautification}

[TOC]

# Introduction {#gapi_fb_intro}

In this tutorial you will learn:
* Basics of a sample face beautification algorithm;
* How to infer different networks inside a pipeline with G-API;
* How to run a G-API pipeline on a video stream.

## Prerequisites {#gapi_fb_prerec}

This sample requires:
- PC with GNU/Linux or Microsoft Windows (Apple macOS is supported but
  was not tested);
- OpenCV 4.2 or later built with Intel® Distribution of [OpenVINO™
  Toolkit](https://docs.openvinotoolkit.org/) (building with [Intel®
  TBB](https://www.threadingbuildingblocks.org/intel-tbb-tutorial) is
  a plus);
- The following topologies from OpenVINO™ Toolkit [Open Model
  Zoo](https://github.com/opencv/open_model_zoo):
  - `face-detection-adas-0001`;
  - `facial-landmarks-35-adas-0002`.

## Face beautification algorithm {#gapi_fb_algorithm}

We will implement a simple face beautification algorithm using a
combination of modern Deep Learning techniques and traditional
Computer Vision. The general idea behind the algorithm is to make
face skin smoother while preserving face features like eyes or a
mouth contrast. The algorithm identifies parts of the face using a DNN
inference, applies different filters to the parts found, and then
combines it into the final result using basic image arithmetics:

\dot
strict digraph Pipeline {
  node [shape=record fontname=Helvetica fontsize=10 style=filled color="#4c7aa4" fillcolor="#5b9bd5" fontcolor="white"];
  edge [color="#62a8e7"];
  ordering="out";
  splines=ortho;
  rankdir=LR;

  input [label="Input"];
  fd [label="Face\ndetector"];
  bgMask [label="Generate\nBG mask"];
  unshMask [label="Unsharp\nmask"];
  bilFil [label="Bilateral\nfilter"];
  shMask [label="Generate\nsharp mask"];
  blMask [label="Generate\nblur mask"];
  mul_1 [label="*" fontsize=24 shape=circle labelloc=b];
  mul_2 [label="*" fontsize=24 shape=circle labelloc=b];
  mul_3 [label="*" fontsize=24 shape=circle labelloc=b];

  subgraph cluster_0 {
    style=dashed
    fontsize=10
    ld [label="Landmarks\ndetector"];
    label="for each face"
  }

  sum_1 [label="+" fontsize=24 shape=circle];
  out [label="Output"];

  temp_1 [style=invis shape=point width=0];
  temp_2 [style=invis shape=point width=0];
  temp_3 [style=invis shape=point width=0];
  temp_4 [style=invis shape=point width=0];
  temp_5 [style=invis shape=point width=0];
  temp_6 [style=invis shape=point width=0];
  temp_7 [style=invis shape=point width=0];
  temp_8 [style=invis shape=point width=0];
  temp_9 [style=invis shape=point width=0];

  input -> temp_1 [arrowhead=none]
  temp_1 -> fd -> ld
  ld -> temp_4 [arrowhead=none]
  temp_4 -> bgMask
  bgMask -> mul_1 -> sum_1 -> out

  temp_4 -> temp_5 -> temp_6 [arrowhead=none constraint=none]
  ld -> temp_2 -> temp_3 [style=invis constraint=none]

  temp_1 -> {unshMask, bilFil}
  fd -> unshMask [style=invis constraint=none]
  unshMask -> bilFil [style=invis constraint=none]

  bgMask -> shMask [style=invis constraint=none]
  shMask -> blMask [style=invis constraint=none]
  mul_1 -> mul_2 [style=invis constraint=none]
  temp_5 -> shMask -> mul_2
  temp_6 -> blMask -> mul_3

  unshMask -> temp_2 -> temp_5 [style=invis]
  bilFil -> temp_3 -> temp_6 [style=invis]

  mul_2 -> temp_7 [arrowhead=none]
  mul_3 -> temp_8 [arrowhead=none]

  temp_8 -> temp_7 [arrowhead=none constraint=none]
  temp_7 -> sum_1 [constraint=none]

  unshMask -> mul_2 [constraint=none]
  bilFil -> mul_3 [constraint=none]
  temp_1 -> mul_1 [constraint=none]
}
\enddot

Briefly the algorithm is described as follows:
- Input image \f$I\f$ is passed to unsharp mask and bilateral filters
  (\f$U\f$ and \f$L\f$ respectively);
- Input image \f$I\f$ is passed to an SSD-based face detector;
- SSD result (a \f$[1 \times 1 \times 200 \times 7]\f$ blob) is parsed
  and converted to an array of faces;
- Every face is passed to a landmarks detector;
- Based on landmarks found for every face, three image masks are
  generated:
  - A background mask \f$b\f$ -- indicating which areas from the
    original image to keep as-is;
  - A face part mask \f$p\f$ -- identifying regions to preserve
    (sharpen).
  - A face skin mask \f$s\f$ -- identifying regions to blur;
- The final result \f$O\f$ is a composition of features above
  calculated as \f$O = b*I + p*U + s*L\f$.

Generating face element masks based on a limited set of features (just
35 per face, including all its parts) is not very trivial and is
described in the sections below.

# Constructing a G-API pipeline {#gapi_fb_pipeline}

## Declaring Deep Learning topologies {#gapi_fb_decl_nets}

This sample is using two DNN detectors. Every network takes one input
and produces one output. In G-API, networks are defined with macro
G_API_NET():

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp net_decl

To get more information, see
[Declaring Deep Learning topologies](@ref gapi_ifd_declaring_nets)
described in the "Face Analytics pipeline" tutorial.

## Describing the processing graph {#gapi_fb_ppline}

The code below generates a graph for the algorithm above:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp ppl

The resulting graph is a mixture of G-API's standard operations,
user-defined operations (namespace `custom::`), and DNN inference.
The generic function `cv::gapi::infer<>()` allows to trigger inference
within the pipeline; networks to infer are specified as template
parameters.  The sample code is using two versions of `cv::gapi::infer<>()`:
- A frame-oriented one is used to detect faces on the input frame.
- An ROI-list oriented one is used to run landmarks inference on a
  list of faces -- this version produces an array of landmarks per
  every face.

More on this in "Face Analytics pipeline"
([Building a GComputation](@ref gapi_ifd_gcomputation) section).

## Unsharp mask in G-API {#gapi_fb_unsh}

The unsharp mask \f$U\f$ for image \f$I\f$ is defined as:

\f[U = I - s * L(M(I)),\f]

where \f$M()\f$ is a median filter, \f$L()\f$ is the Laplace operator,
and \f$s\f$ is a strength coefficient. While G-API doesn't provide
this function out-of-the-box, it is expressed naturally with the
existing G-API operations:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp unsh

Note that the code snipped above is a regular C++ function defined
with G-API types. Users can write functions like this to simplify
graph construction; when called, this function just puts the relevant
nodes to the pipeline it is used in.

# Custom operations {#gapi_fb_proc}

The face beautification graph is using custom operations
extensively. This chapter focuses on the most interesting kernels,
refer to [G-API Kernel API](@ref gapi_kernel_api) for general
information on defining operations and implementing kernels in G-API.

## Face detector post-processing {#gapi_fb_face_detect}

A face detector output is converted to an array of faces with the
following kernel:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp vec_ROI
@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp fd_pp

## Facial landmarks post-processing {#gapi_fb_landm_detect}

The algorithm infers locations of face elements (like the eyes, the mouth
and the head contour itself) using a generic facial landmarks detector
(<a href="https://github.com/opencv/open_model_zoo/blob/master/models/intel/facial-landmarks-35-adas-0002/description/facial-landmarks-35-adas-0002.md">details</a>)
from OpenVINO™ Open Model Zoo. However, the detected landmarks as-is are not
enough to generate masks --- this operation requires regions of interest on
the face represented by closed contours, so some interpolation is applied to
get them. This landmarks
processing and interpolation is performed by the following kernel:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp ld_pp_cnts

The kernel takes two arrays of denormalized landmarks coordinates and
returns an array of elements' closed contours and an array of faces'
closed contours; in other words, outputs are, the first, an array of
contours of image areas to be sharpened and, the second, another one
to be smoothed.

Here and below `Contour` is a vector of points.

### Getting an eye contour {#gapi_fb_ld_eye}

Eye contours are estimated with the following function:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp ld_pp_incl
@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp ld_pp_eye

Briefly, this function restores the bottom side of an eye by a
half-ellipse based on two points in left and right eye
corners. In fact, `cv::ellipse2Poly()` is used to approximate the eye region, and
the function only defines ellipse parameters based on just two points:
- The ellipse center and the \f$X\f$ half-axis calculated by two eye Points;
- The \f$Y\f$ half-axis calculated according to the assumption that an average
eye width is \f$1/3\f$ of its length;
- The start and the end angles which are 0 and 180 (refer to
  `cv::ellipse()` documentation);
- The angle delta: how much points to produce in the contour;
- The inclination angle of the axes.

The use of the `atan2()` instead of just `atan()` in function
`custom::getLineInclinationAngleDegrees()` is essential as it allows to
return a negative value depending on the `x` and the `y` signs so we
can get the right angle even in case of upside-down face arrangement
(if we put the points in the right order, of course).

### Getting a forehead contour {#gapi_fb_ld_fhd}

The function  approximates the forehead contour:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp ld_pp_fhd

As we have only jaw points in our detected landmarks, we have to get a
half-ellipse based on three points of a jaw: the leftmost, the
rightmost and the lowest one. The jaw width is assumed to be equal to the
forehead width and the latter is calculated using the left and the
right points. Speaking of the \f$Y\f$ axis, we have no points to get
it directly, and instead assume that the forehead height is about \f$2/3\f$
of the jaw height, which can be figured out from the face center (the
middle between the left and right points) and the lowest jaw point.

## Drawing masks {#gapi_fb_masks_drw}

When we have all the contours needed, we are able to draw masks:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp msk_ppline

The steps to get the masks are:
* the "sharp" mask calculation:
    * fill the contours that should be sharpened;
    * blur that to get the "sharp" mask (`mskSharpG`);
* the "bilateral" mask calculation:
    * fill all the face contours fully;
    * blur that;
    * subtract areas which intersect with the "sharp" mask --- and get the
      "bilateral" mask (`mskBlurFinal`);
* the background mask calculation:
    * add two previous masks
    * set all non-zero pixels of the result as 255 (by `cv::gapi::threshold()`)
    * revert the output (by `cv::gapi::bitwise_not`) to get the background
      mask (`mskNoFaces`).

# Configuring and running the pipeline {#gapi_fb_comp_args}

Once the graph is fully expressed, we can finally compile it and run
on real data. G-API graph compilation is the stage where the G-API
framework actually understands which kernels and networks to use. This
configuration happens via G-API compilation arguments.

## DNN parameters {#gapi_fb_comp_args_net}

This sample is using OpenVINO™ Toolkit Inference Engine backend for DL
inference, which is configured the following way:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp net_param

Every `cv::gapi::ie::Params<>` object is related to the network
specified in its template argument. We should pass there the network
type we have defined in `G_API_NET()` in the early beginning of the
tutorial.

Network parameters are then wrapped in `cv::gapi::NetworkPackage`:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp netw

More details in "Face Analytics Pipeline"
([Configuring the pipeline](@ref gapi_ifd_configuration) section).

## Kernel packages  {#gapi_fb_comp_args_kernels}

In this example we use a lot of custom kernels, in addition to that we
use Fluid backend to optimize out memory for G-API's standard kernels
where applicable. The resulting kernel package is formed like this:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp kern_pass_1

## Compiling the streaming pipeline  {#gapi_fb_compiling}

G-API optimizes execution for video streams when compiled in the
"Streaming" mode.

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp str_comp

More on this in "Face Analytics Pipeline"
([Configuring the pipeline](@ref gapi_ifd_configuration) section).

## Running the streaming pipeline {#gapi_fb_running}

In order to run the G-API streaming pipeline, all we need is to
specify the input video source, call
`cv::GStreamingCompiled::start()`, and then fetch the pipeline
processing results:

@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp str_src
@snippet cpp/tutorial_code/gapi/face_beautification/face_beautification.cpp str_loop

Once results are ready and can be pulled from the pipeline we display
it on the screen and handle GUI events.

See [Running the pipeline](@ref gapi_ifd_running) section
in the "Face Analytics Pipeline" tutorial for more details.

# Conclusion {#gapi_fb_cncl}

The tutorial has two goals: to show the use of brand new features of
G-API introduced in OpenCV 4.2, and give a basic understanding on a
sample face beautification algorithm.

The result of the algorithm application:

![Face Beautification example](pics/example.jpg)

On the test machine (Intel® Core™ i7-8700) the G-API-optimized video
pipeline outperforms its serial (non-pipelined) version by a factor of
**2.7** -- meaning that for such a non-trivial graph, the proper
pipelining can bring almost 3x increase in performance.

<!---
The idea in general is to implement a real-time video stream processing that
detects faces and applies some filters to make them look beautiful (more or
less). The pipeline is the following:

Two topologies from OMZ have been used in this sample: the
<a href="https://github.com/opencv/open_model_zoo/tree/master/models/intel
/face-detection-adas-0001">face-detection-adas-0001</a>
and the
<a href="https://github.com/opencv/open_model_zoo/blob/master/models/intel
/facial-landmarks-35-adas-0002/description/facial-landmarks-35-adas-0002.md">
facial-landmarks-35-adas-0002</a>.

The face detector takes the input image and returns a blob with the shape
[1,1,200,7] after the inference (200 is the maximum number of
faces which can be detected).
In order to process every face individually, we need to convert this output to a
list of regions on the image.

The masks for different filters are built based on facial landmarks, which are
inferred for every face. The result of the inference
is a blob with 35 landmarks: the first 18 of them are facial elements
(eyes, eyebrows, a nose, a mouth) and the last 17 --- a jaw contour. Landmarks
are floating point values of coordinates normalized relatively to an input ROI
(not the original frame). In addition, for the further goals we need contours of
eyes, mouths, faces, etc., not the landmarks. So, post-processing of the Mat is
also required here. The process is split into two parts --- landmarks'
coordinates denormalization to the real pixel coordinates of the source frame
and getting necessary closed contours based on these coordinates.

The last step of processing the inference data is drawing masks using the
calculated contours. In this demo the contours don't need to be pixel accurate,
since masks are blurred with Gaussian filter anyway. Another point that should
be mentioned here is getting
three masks (for areas to be smoothed, for ones to be sharpened and for the
background) which have no intersections with each other; this approach allows to
apply the calculated masks to the corresponding images prepared beforehand and
then just to summarize them to get the output image without any other actions.

As we can see, this algorithm is appropriate to illustrate G-API usage
convenience and efficiency in the context of solving a real CV/DL problem.

(On detector post-proc)
Some points to be mentioned about this kernel implementation:

- It takes a `cv::Mat` from the detector and a `cv::Mat` from the input; it
returns an array of ROI's where faces have been detected.

- `cv::Mat` data parsing by the pointer on a float is used here.

- By far the most important thing here is solving an issue that sometimes
detector returns coordinates located outside of the image; if we pass such an
ROI to be processed, errors in the landmarks detection will occur. The frame box
`borders` is created and then intersected with the face rectangle
(by `operator&()`) to handle such cases and save the ROI which is for sure
inside the frame.

Data parsing after the facial landmarks detector happens according to the same
scheme with inconsiderable adjustments.


## Possible further improvements

There are some points in the algorithm to be improved.

### Correct ROI reshaping for meeting conditions required by the facial landmarks detector

The input of the facial landmarks detector is a square ROI, but the face
detector gives non-square rectangles in general. If we let the backend within
Inference-API compress the rectangle to a square by itself, the lack of
inference accuracy can be noticed in some cases.
There is a solution: we can give a describing square ROI instead of the
rectangular one to the landmarks detector, so there will be no need to compress
the ROI, which will lead to accuracy improvement.
Unfortunately, another problem occurs if we do that:
if the rectangular ROI is near the border, a describing square will probably go
out of the frame --- that leads to errors of the landmarks detector.
To aviod such a mistake, we have to implement an algorithm that, firstly,
describes every rectangle by a square, then counts the farthest coordinates
turned up to be outside of the frame and, finally, pads the source image by
borders (e.g. single-colored) with the size counted. It will be safe to take
square ROIs for the facial landmarks detector after that frame adjustment.

### Research for the best parameters (used in GaussianBlur() or unsharpMask(), etc.)

### Parameters autoscaling

-->