VideoAnalysis/detection.py at master · jiarongqiu/VideoAnalysis · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Created by charlie on 18-4-24

from __future__ import print_function
from __future__ import absolute_import
from __future__ import division

from data_provider.THUMOS14 import THUMOS14
from sklearn.svm import SVC
from sklearn.metrics import accuracy_score
import math
import numpy as np
import pickle
import os
from collections import Counter
import pandas as pd
from proposal.evaluate import cal_iou
import time

def mAP(ground_truth, proposals, num_proposals=200, tiou=0.5):
    '''
    mean average precision of each category
    :param groud_truth: video,start,end,label
    :param proposals:video,start,end,label
    :return:
    '''
    proposals = pd.DataFrame(proposals)
    proposals = proposals.sort_values(4, ascending=False)
    ground_truth = pd.DataFrame(ground_truth)
    videos = pd.Series.unique(ground_truth.loc[:, 0])
    if len(videos) * num_proposals < len(proposals):
        proposals = proposals.loc[:len(videos) * num_proposals, :]
    print(len(proposals))
    actions = sorted(pd.Series.unique(ground_truth.loc[:, 3]))
    precision=[]
    for action in actions:
        y_pred = proposals.loc[proposals.loc[:, 3] == action]
        y = ground_truth.loc[ground_truth.loc[:, 3] == action]
        y_pred = y_pred.sort_values(4, ascending=False)
        precision_recall = []
        right = fetch = total=0
        cnt=0
        used=set()
        for idx, row in y_pred.iterrows():
            total+=1
            video, start, end, label, _ = row
            cnt+=1
            # if cnt>50:break
            for idx, row in y.loc[y.loc[:, 0] == video].iterrows():
                if idx in used:continue
                _, ground_truth_start, ground_truth_end, ground_truth_label = row
                # print(start,end,ground_truth_start,ground_truth_end,cal_iou((start, end), (ground_truth_start, ground_truth_end)))
                if cal_iou((start, end), (ground_truth_start, ground_truth_end)) > tiou:
                    # print(idx,ground_truth_label,label,cal_iou((start, end), (ground_truth_start, ground_truth_end)))
                    fetch += 1
                    used.add(idx)
                    if ground_truth_label == label:
                        right += 1
                    break
            if fetch == 0:
                precision_recall.append((0, 0))
            else:
                precision_recall.append((right / total, fetch))
        precision_recall=pd.DataFrame(precision_recall)
        grouped=precision_recall.loc[:,0].groupby(precision_recall.loc[:,1])
        grouped_precision=grouped.max()
        precision.append(sum(grouped_precision)/len(grouped_precision))
        # break
    print(sum(precision)/len(precision),precision)


def train():
    dataset = THUMOS14()
    train, test = dataset.load_in_info()
    label2idx = dataset.load_in_label()
    X, y = [], []
    cnt = 0
    for video, info in train.iteritems():
        proposals = info['proposals']
        for p in proposals:
            start = math.floor(p['start'])
            end = math.floor(p['end'])
            label = p['label']
            idx = label2idx[label]
            while start <= end:
                feature = dataset.load_feature(video, start, start + 1)
                X.append(feature)
                y.append(idx)
                start += 1
        # cnt += 1
        # if cnt > 10: break
    X = np.array(X)
    y = np.array(y)
    clf = SVC(kernel='linear')
    clf.fit(X=X, y=y)
    pred = clf.predict(X)
    print(X.shape, y.shape)
    print(accuracy_score(y_true=y, y_pred=pred))
    # pickle.dump(clf, open(os.path.join(THUMOS14.MODEL_DIR, 'detection_svm'), 'w'))


def test():
    clf = pickle.load(open(os.path.join(THUMOS14.MODEL_DIR, 'detection_svm'), 'r'))
    dataset = THUMOS14()
    _, test = dataset.load_in_info()
    label2idx = dataset.load_in_label()

    ground_truth, proposals = [], []
    result = pickle.load(open(os.path.join(THUMOS14.RES_DIR, 'turn_proposal'), 'r'))
    for video, info in result.iteritems():
        candidates = info['proposals']
        print(video, len(candidates))
        for proposal in candidates:
            start = math.floor(proposal['start'])
            end = math.floor(proposal['end'])
            score = proposal['score']
            if start>end:continue
            pt=start
            X = []
            while pt <= end:
                feature = dataset.load_feature(video, pt, pt + 1)
                X.append(feature)
                pt += 1
            X = np.array(X)
            pred = clf.predict(X)
            most_common, _ = Counter(pred).most_common(1)[0]
            proposal['label'] = most_common
            proposals.append((video, start, end, most_common, score))
    for video, info in test.iteritems():
        for p in info['proposals']:
            start = p['start']
            end = p['end']
            label = p['label']
            idx = label2idx[label]
            ground_truth.append((video, start, end, idx))
    pickle.dump([ground_truth, proposals], open(os.path.join(THUMOS14.RES_DIR, 'detection_result'), 'w'))


if __name__ == '__main__':
    # train()
    # test()
    ground_truth, proposals = pickle.load(open(os.path.join(THUMOS14.RES_DIR, 'detection_result'), 'r'))
    mAP(ground_truth, proposals)