# -*- coding: utf-8 -*-
"""
Evaluate results of 2D face tracking based on particle filter algorithm.
@author: Bessie Domínguez-Dáger
"""
import numpy as np
from pftracker.modules.metrics.PrecRecall import calc_PrecRecall
[docs]def error(det_tuple, gt_file):
"""2D face tracking evaluation.
This function reads ground truth annotations of bounding box from .txt
files and then calculates precision, recall and F1-score error metrics
between detections that have been obtained from particle filter estimation
and ground truth annotations.
Ground truth annotations shoud be in the format:
- x,y,w,h
where (x, y) are the center bounding box coordinates and (w, h) are the
bounding box width and height respectively.
Besides, it is possible to read annotations from the 'YouTube Dataset'
format:
- name, [ignore], x, y, w, h, [ignore], [ignore]
This dataset was used for testing the framework and is available at
www.cs.tau.ac.il/˜wolf/ytfaces
Args:
det_tuple(tuple): tuple with particle filter estimation results
gt_file(str): .txt file directory of ground truth annotations
Returns:
8-element tuple containing
- **P** (*array*): precision value per frame, array with
(1, number_of_frames) dimension
- **R** (*array*): recall value per frame, array with
(1, number_of_frames) dimension
- **P_mean** (*float*): precision mean value
- **R_mean** (*float*): recall mean value
- **P_std** (*float*): precision standard deviation value
- **R_std** (*float*): recall standard deviation value
- **F1Score_mean** (*float*): F-1-score mean metric
- **F1Score_std** (*float*): F-1-score standard deviation value
"""
det = np.zeros((len(det_tuple), 4)).astype(int)
det_temp = np.zeros((len(det_tuple), 4))
for j in range(len(det_tuple)):
det_temp[j] = det_tuple[j]
det[j, 0:2] = det_temp[j, 0:2] - det_temp[j, 2:]//2
det[j, 2:] = det_temp[j, 0:2] + det_temp[j, 2:]//2
gtFile = open(gt_file)
gt_list = gtFile.readlines()
gt = np.zeros((len(det_tuple), 4)).astype(int)
gt_temp = np.zeros((len(det_tuple), 4)).astype(int)
for i in range(len(det_tuple)):
if len(gt_list[i].split(",")) > 4:
gt_temp[i] = gt_list[i].split(",")[2:-2]
elif len(gt_list[i].split(",")) == 4:
gt_temp[i]= gt_list[i].split(",")
gt[i, 0:2] = gt_temp[i, 0:2] - gt_temp[i, 2:]//2
gt[i, 2:] = gt_temp[i, 0:2] + gt_temp[i, 2:]//2
gtFile.close()
# compute precision, recall and F1-score metrics
P, R, F1Score = calc_PrecRecall(det, gt)
# print(type(P[0]))
# P = P.astype(float)
# R = R.astype(float)
# print(type(P[0]))
# print(P[0])
# precision and recall average
P_mean = np.sum(P)/len(P)
R_mean = np.sum(R)/len(R)
# precision and recall standard deviation
P_std = np.std(P, dtype=np.float64)
R_std = np.std(R, dtype=np.float64)
# F1-score average
F1Score_mean = np.sum(F1Score)/len(F1Score)
# precision and recall standard deviation
F1Score_std = np.std(F1Score, dtype=np.float64)
return P, R, P_mean, R_mean, P_std, R_std, F1Score_mean, F1Score_std