shapepredictor.h

Go to the documentation of this file.

/* ============================================================
 *
 * This file is a part of digiKam project
 * https://www.digikam.org
 *
 * Date        : 16/08/2016
 * Description : A Shape predictor class that can predicts 68
 *               facial point including points surrounding faces
 *               eyes, that can be used for detecting human eyes
 *               positions, almost all codes are ported from dlib
 *               library (dlib.net/)
 *
 * Copyright (C) 2016      by Omar Amin <Omar dot moh dot amin at gmail dot com>
 * Copyright (C) 2019      by Thanh Trung Dinh <dinhthanhtrung1996 at gmail dot com>
 * Copyright (C) 2016-2022 by Gilles Caulier <caulier dot gilles at gmail dot com>
 *
 * This program is free software; you can redistribute it
 * and/or modify it under the terms of the GNU General
 * Public License as published by the Free Software Foundation;
 * either version 2, or (at your option)
 * any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * ============================================================ */
 
#ifndef DIGIKAM_SHAPE_PREDICTOR_H
#define DIGIKAM_SHAPE_PREDICTOR_H
 
// C++ includes
 
#include <vector>
 
// Local includes
 
#include "digikam_opencv.h"
#include "pointtransformaffine.h"
#include "vectoroperations.h"
#include "matrixoperations.h"
#include "fullobjectdetection.h"
#include "qdatastreamoverloads.h"
 
namespace Digikam
{
 
namespace RedEye
{
 
struct SplitFeature
{
    unsigned long idx1;
    unsigned long idx2;
    float         thresh;
};
 
QDataStream& operator << (QDataStream& dataStream, const SplitFeature& sp);
QDataStream& operator >> (QDataStream& dataStream, SplitFeature& sp);
 
// NOTE: a tree is just a std::vector<RedEye::SplitFeature>. We use this function to navigate the tree nodes.
 
unsigned long left_child(unsigned long idx);
 
unsigned long right_child(unsigned long idx);
 
// ----------------------------------------------------------------------------------------
 
struct RegressionTree
{
    std::vector<SplitFeature>        splits;
    std::vector<std::vector<float> > leaf_values;
 
    unsigned long num_leaves() const;
 
    const std::vector<float>& operator()(const std::vector<float>& feature_pixel_values,
                                         unsigned long& i) const;
};
 
QDataStream& operator << (QDataStream& dataStream, const RegressionTree& regtree);
QDataStream& operator >> (QDataStream& dataStream, RegressionTree& regtree);
 
template<class T>
inline std::vector<T> location(const std::vector<T>& shape,
                               unsigned long idx)
{
    std::vector<T> temp(2);
    temp[0] = shape[idx * 2    ];
    temp[1] = shape[idx * 2 + 1];
 
    return temp;
}
 
// ------------------------------------------------------------------------------------
 
unsigned long nearestShapePoint(const std::vector<float>& shape,
                                const std::vector<float>& pt);
 
// ------------------------------------------------------------------------------------
 
void createShapeRelativeEncoding(const std::vector<float>& shape,
                                 const std::vector<std::vector<float> >& pixel_coordinates,
                                 std::vector<unsigned long>& anchor_idx,
                                 std::vector<std::vector<float> >& deltas);
 
// ------------------------------------------------------------------------------------
 
PointTransformAffine findTformBetweenShapes(const std::vector<float>& from_shape,
                                            const std::vector<float>& to_shape);
 
// ------------------------------------------------------------------------------------
 
PointTransformAffine normalizingTform(const cv::Rect& rect);
 
// ------------------------------------------------------------------------------------
 
PointTransformAffine unnormalizingTform(const cv::Rect& rect);
bool pointContained(const cv::Rect& rect, const std::vector<float>& point);
 
// ------------------------------------------------------------------------------------
 
void extractFeaturePixelValues(const cv::Mat& img_,
                               const cv::Rect& rect,
                               const std::vector<float>& current_shape,
                               const std::vector<float>& reference_shape,
                               const std::vector<unsigned long>& reference_pixel_anchor_idx,
                               const std::vector<std::vector<float> >& reference_pixel_deltas,
                               std::vector<float>& feature_pixel_values);
 
// ------------------------------------------------------------------------------------
 
class ShapePredictor
{
public:
 
    explicit ShapePredictor();
 
    unsigned long num_parts()                            const;
    unsigned long num_features()                         const;
 
    FullObjectDetection operator()(const cv::Mat& img,
                                   const cv::Rect& rect) const;
 
public:
 
    std::vector<float>                                initial_shape;
    std::vector<std::vector<RedEye::RegressionTree> > forests;
    std::vector<std::vector<unsigned long> >          anchor_idx;
    std::vector<std::vector<std::vector<float> > >    deltas;
};
 
QDataStream& operator << (QDataStream& dataStream, const ShapePredictor& shape);
QDataStream& operator >> (QDataStream& dataStream, ShapePredictor& shape);
 
} // namespace RedEye
 
} // namespace Digikam
 
#endif // DIGIKAM_SHAPE_PREDICTOR_H