webrtc/rtc_tools/frame_analyzer/linear_least_squares.h
Magnus Jedvert 1927dfafab Add tool for aligning color space of video files
This class adds logic for aligning color space of a test video compared
to a reference video. If there is a color space mismatch, it typically
does not have much impact on human perception, but it has a big impact
on PSNR and SSIM calculations. For example, aligning a test run with VP8
improves PSNR and SSIM from:
Average PSNR: 29.142818, average SSIM: 0.946026
to:
Average PSNR: 38.146229, average SSIM: 0.965388.

The optiomal color transformation between the two videos were:
0.86 0.01 0.00 14.37
0.00 0.88 0.00 15.32
0.00 0.00 0.88 15.74
which is converting YUV full range to YUV limited range. There is
already a CL out for fixing this discrepancy here:
https://webrtc-review.googlesource.com/c/src/+/94543

After that, hopefully there is no color space mismatch when saving the
raw YUV values. It's good that the video quality tool is color space
agnostic anyway, and can compensate for differences when the test
video is obtained by e.g. filming a physical device screen.

Also, the linear least square logic will be used for compensating
geometric distorisions in a follow-up CL.

Bug: webrtc:9642
Change-Id: I499713960a0544d8e45c5d09886e68ec829b28a7
Reviewed-on: https://webrtc-review.googlesource.com/c/95950
Reviewed-by: Sami Kalliomäki <sakal@webrtc.org>
Reviewed-by: Patrik Höglund <phoglund@webrtc.org>
Commit-Queue: Magnus Jedvert <magjed@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#25193}
2018-10-16 07:55:37 +00:00

53 lines
2 KiB
C++

/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef RTC_TOOLS_FRAME_ANALYZER_LINEAR_LEAST_SQUARES_H_
#define RTC_TOOLS_FRAME_ANALYZER_LINEAR_LEAST_SQUARES_H_
#include <valarray>
#include <vector>
#include "absl/types/optional.h"
namespace webrtc {
namespace test {
// This class is used for finding a matrix b that roughly solves the equation:
// y = x * b. This is generally impossible to do exactly, so the problem is
// rephrased as finding the matrix b that minimizes the difference:
// |y - x * b|^2. Calling multiple AddObservations() is equivalent to
// concatenating the observation vectors and calling AddObservations() once. The
// reason for doing it incrementally is that we can't store the raw YUV values
// for a whole video file in memory at once. This class has a constant memory
// footprint, regardless how may times AddObservations() is called.
class IncrementalLinearLeastSquares {
public:
IncrementalLinearLeastSquares();
~IncrementalLinearLeastSquares();
// Add a number of observations. The subvectors of x and y must have the same
// length.
void AddObservations(const std::vector<std::vector<uint8_t>>& x,
const std::vector<std::vector<uint8_t>>& y);
// Calculate and return the best linear solution, given the observations so
// far.
std::vector<std::vector<double>> GetBestSolution() const;
private:
// Running sum of x^T * x.
absl::optional<std::valarray<std::valarray<uint64_t>>> sum_xx;
// Running sum of x^T * y.
absl::optional<std::valarray<std::valarray<uint64_t>>> sum_xy;
};
} // namespace test
} // namespace webrtc
#endif // RTC_TOOLS_FRAME_ANALYZER_LINEAR_LEAST_SQUARES_H_