blob: 6e8ab9052afbd23e8cc0fcfc2dcd3cc4f971834c [file] [log] [blame]
#!/usr/bin/env python
## Copyright (c) 2021, Alliance for Open Media. All rights reserved
##
## This source code is subject to the terms of the BSD 3-Clause Clear License and the
## Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear License was
## not distributed with this source code in the LICENSE file, you can obtain it
## at aomedia.org/license/software-license/bsd-3-c-c/. If the Alliance for Open Media Patent
## License 1.0 was not distributed with this source code in the PATENTS file, you
## can obtain it at aomedia.org/license/patent-license/.
##
__author__ = "maggie.sun@intel.com, ryanlei@fb.com"
import numpy as np
import math
import scipy.interpolate
import logging
from Config import LoggerName
from operator import itemgetter
subloggername = "CalcBDRate"
loggername = LoggerName + '.' + '%s' % subloggername
logger = logging.getLogger(loggername)
def non_decreasing(L):
return all(x<=y for x, y in zip(L, L[1:]))
def check_monotonicity(RDPoints):
'''
check if the input list of RD points are monotonic, assuming the input
has been sorted in the quality value non-decreasing order. expect the bit
rate should also be in the non-decreasing order
'''
br = [RDPoints[i][0] for i in range(len(RDPoints))]
qty = [RDPoints[i][1] for i in range(len(RDPoints))]
return non_decreasing(br) and non_decreasing(qty)
def filter_vmaf_non_monotonic(br_qty_pairs):
'''
To solve the problem with VMAF non-monotonicity in a flat (saturated)
region of the curve, if VMAF non-monotonicity happens at VMAF value
99.5 or above, the non-monotonic value and the values corresponding
to bitrates higher than the non-monotonic value are excluded from the
BD-rate calculation. The VMAF BD-rate number is still reported and
used in the VMAF metric average.
'''
#first sort input RD pairs by bit rate
out_br_qty_pairs = []
br_qty_pairs.sort(key = itemgetter(0, 1))
for i in range(len(br_qty_pairs)):
if (i != 0 and
br_qty_pairs[i][0] >= out_br_qty_pairs[-1][0] and
br_qty_pairs[i][1] < out_br_qty_pairs[-1][1] and
out_br_qty_pairs[-1][1] >= 99.5):
break
else:
out_br_qty_pairs.append(br_qty_pairs[i])
return out_br_qty_pairs
# BJONTEGAARD Bjontegaard metric
# Calculation is adapted from Google implementation
# PCHIP method - Piecewise Cubic Hermite Interpolating Polynomial interpolation
def BD_RATE(qty_type, br1, qtyMtrc1, br2, qtyMtrc2):
brqtypairs1 = []; brqtypairs2 = []
for i in range(min(len(qtyMtrc1), len(br1))):
if (br1[i] != '' and qtyMtrc1[i] != ''):
brqtypairs1.append((br1[i], qtyMtrc1[i]))
for i in range(min(len(qtyMtrc2), len(br2))):
if (br2[i] != '' and qtyMtrc2[i] != ''):
brqtypairs2.append((br2[i], qtyMtrc2[i]))
if (qty_type == 'VMAF_Y' or qty_type == 'VMAF_Y-NEG'):
brqtypairs1 = filter_vmaf_non_monotonic(brqtypairs1)
brqtypairs2 = filter_vmaf_non_monotonic(brqtypairs2)
# sort the pair based on quality metric values in increasing order
# if quality metric values are the same, then sort the bit rate in increasing order
brqtypairs1.sort(key = itemgetter(1, 0))
brqtypairs2.sort(key = itemgetter(1, 0))
rd1_monotonic = check_monotonicity(brqtypairs1)
rd2_monotonic = check_monotonicity(brqtypairs2)
if (rd1_monotonic == False or rd2_monotonic == False):
return (-1, "Error: Non-monotonic")
try:
logbr1 = [math.log(x[0]) for x in brqtypairs1]
qmetrics1 = [100.0 if x[1] == float('inf') else x[1] for x in brqtypairs1]
logbr2 = [math.log(x[0]) for x in brqtypairs2]
qmetrics2 = [100.0 if x[1] == float('inf') else x[1] for x in brqtypairs2]
except ValueError:
return (-1, "Error: Invalid Input Data")
if not brqtypairs1 or not brqtypairs2:
logger.info("Error: one of input lists is empty!")
return (-1, "Error: one of input lists is empty!")
# remove duplicated quality metric value, the RD point with higher bit rate is removed
dup_idx = [i for i in range(1, len(qmetrics1)) if qmetrics1[i - 1] == qmetrics1[i]]
for idx in sorted(dup_idx, reverse=True):
del qmetrics1[idx]
del logbr1[idx]
dup_idx = [i for i in range(1, len(qmetrics2)) if qmetrics2[i - 1] == qmetrics2[i]]
for idx in sorted(dup_idx, reverse=True):
del qmetrics2[idx]
del logbr2[idx]
# find max and min of quality metrics
min_int = max(min(qmetrics1), min(qmetrics2))
max_int = min(max(qmetrics1), max(qmetrics2))
if min_int >= max_int:
logger.info("Error: no overlap from input 2 lists of quality metrics!")
return (-1, "Error: no overlap from input 2 lists of quality metrics!")
# generate samples between max and min of quality metrics
lin = np.linspace(min_int, max_int, num=100, retstep=True)
interval = lin[1]
samples = lin[0]
# interpolation
v1 = scipy.interpolate.pchip_interpolate(qmetrics1, logbr1, samples)
v2 = scipy.interpolate.pchip_interpolate(qmetrics2, logbr2, samples)
# Calculate the integral using the trapezoid method on the samples.
int1 = np.trapz(v1, dx=interval)
int2 = np.trapz(v2, dx=interval)
# find avg diff
avg_exp_diff = (int2 - int1) / (max_int - min_int)
avg_diff = (math.exp(avg_exp_diff) - 1) * 100
return (0, round(avg_diff, 4))
'''
if __name__ == "__main__":
br1 = [9563.04, 6923.28, 4894.8, 3304.32, 2108.4, 1299.84]
#qty1 = [50.0198, 46.9709, 43.4791, 39.6659, 35.8063, 32.3055]
#qty1 = [50.0198, 46.9709, 43.4791, 48.0000, 35.8063, 32.3055]
qty1 = [99.8198, 99.7709, 98.4791, 99.5000, 98.8063, 98.3055]
br2 = [9758.88, 7111.68, 5073.36, 3446.4, 2178, 1306.56]
#qty2 = [49.6767, 46.7027, 43.2038, 39.297, 35.2944, 31.5938]
qty2 = [99.8767, 99.7027, 99.2038, 99.200, 98.2944, 97.5938]
qty_type = 'VMAF-Y'
plot_rd_curve(br1, qty1, qty_type, 'r', '-', 'o')
plot_rd_curve(br2, qty2, qty_type, 'b', '-', '*')
plt.show()
bdrate = BD_RATE('VMAF_Y', br1, qty1, br2, qty2)
if bdrate != 'Non-monotonic Error':
print("bdrate calculated is %3.3f%%" % bdrate)
else:
print("there is Non-monotonic Error in bdrate calculation")
'''