BiSCC/proposal_methods_eval.py at master · lgzlIlIlI/BiSCC · 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
146
147
import numpy as np
import torch
import utils.wsad_utils as utils
from scipy.signal import savgol_filter
import pandas as pd
import options
args = options.parser.parse_args()

def filter_segments(segment_predict, vn):
    ambilist = './Thumos14reduced-Annotations/Ambiguous_test.txt'
    try:
        ambilist = list(open(ambilist, "r"))
        ambilist = [a.strip("\n").split(" ") for a in ambilist]
    except:
        ambilist = []
    ind = np.zeros(np.shape(segment_predict)[0])
    for i in range(np.shape(segment_predict)[0]):
        #s[j], e[j], np.max(seg)+0.7*c_s[c],c]
        for a in ambilist:
            if a[0] == vn:
                gt = range(
                    int(round(float(a[2]) * 25 / 16)), int(round(float(a[3]) * 25 / 16))
                )
                pd = range(int(segment_predict[i][0]), int(segment_predict[i][1]))
                IoU = float(len(set(gt).intersection(set(pd)))) / float(
                    len(set(gt).union(set(pd)))
                )
                if IoU > 0:
                    ind[i] = 1
    s = [
        segment_predict[i, :]
        for i in range(np.shape(segment_predict)[0])
        if ind[i] == 0
    ]
    return np.array(s)

def filter_segments_2(segment_predict, vn):
    ambilist = './Thumos14reduced-Annotations/Ambiguous_test.txt'
    try:
        ambilist = list(open(ambilist, "r"))
        ambilist = [a.strip("\n").split(" ") for a in ambilist]
    except:
        ambilist = []
    ind = np.zeros(np.shape(segment_predict)[0])
    for i in range(np.shape(segment_predict)[0]):
        #s[j], e[j], np.max(seg)+0.7*c_s[c],c]
        for a in ambilist:
            if a[0] == vn:
                if not (max(float(a[2]),segment_predict[i][0])>=min(float(a[3]),segment_predict[i][1])):
                    ind[i] = 1
    s = [
        segment_predict[i, :]
        for i in range(np.shape(segment_predict)[0])
        if ind[i] == 0
    ]
    return np.array(s)

def smooth(v, order=2,lens=200):
    l = min(lens, len(v))
    l = l - (1 - l % 2)
    if len(v) <= order:
        return v
    return savgol_filter(v, l, order)

def get_topk_mean(x, k, axis=0):
    return np.mean(np.sort(x, axis=axis)[-int(k):, :], axis=0)

def get_cls_score(element_cls, dim=-2, rat=20, ind=None):

    topk_val, _ = torch.topk(element_cls,
                             k=max(1, int(element_cls.shape[-2] // rat)),
                             dim=-2)
    instance_logits = torch.mean(topk_val, dim=-2)
    pred_vid_score = torch.softmax(
        instance_logits, dim=-1)[..., :-1].squeeze().data.cpu().numpy()
    return pred_vid_score

@torch.no_grad()
def multiple_threshold_hamnet(vid_name,data_dict):
    elem = data_dict['cas']
    element_atn=data_dict['atn']
    element_logits = elem * element_atn
    pred_vid_score = get_cls_score(element_logits, rat=10)
    cas_supp = element_logits[..., :-1]
    cas_supp_atn = element_atn
    pred = np.where(pred_vid_score >= 0.2)[0]

    act_thresh = np.linspace(0.1,0.9,10)
    prediction = None
    if len(pred) == 0:
        pred = np.array([np.argmax(pred_vid_score)])
    cas_pred = cas_supp[0].cpu().numpy()[:, pred]
    num_segments = cas_pred.shape[0]
    cas_pred = np.reshape(cas_pred, (num_segments, -1, 1))
    cas_pred_atn = cas_supp_atn[0].cpu().numpy()[:, [0]]
    cas_pred_atn = np.reshape(cas_pred_atn, (num_segments, -1, 1))
    proposal_dict = {}
    for i in range(len(act_thresh)):
        cas_temp = cas_pred.copy()
        cas_temp_atn = cas_pred_atn.copy()
        seg_list = []
        for c in range(len(pred)):
            pos = np.where(cas_temp_atn[:, 0, 0] > act_thresh[i])
            seg_list.append(pos)

        proposals = utils.get_proposal_oic_2(seg_list,cas_temp, pred_vid_score,pred,  args.scale,
                                            num_segments, args.feature_fps,num_segments,gamma=args.gamma_oic)
        for j in range(len(proposals)):
            class_id = proposals[j][0][0]
            if class_id not in proposal_dict.keys():
                proposal_dict[class_id] = []
            proposal_dict[class_id] += proposals[j]

    final_proposals = []
    for class_id in proposal_dict.keys():
        final_proposals.append(
            utils.soft_nms(proposal_dict[class_id], 0.7, sigma=0.3))

    video_lst, t_start_lst, t_end_lst = [], [], []
    label_lst, score_lst = [], []
    segment_predict = []
    for i in range(len(final_proposals)):
        for j in range(len(final_proposals[i])):
            [c_pred, c_score, t_start, t_end] = final_proposals[i][j]
            segment_predict.append([t_start, t_end,c_score,c_pred])

    segment_predict = np.array(segment_predict)
    segment_predict = filter_segments(segment_predict, vid_name.decode())

    video_lst, t_start_lst, t_end_lst = [], [], []
    label_lst, score_lst = [], []
    for i in range(np.shape(segment_predict)[0]):
        video_lst.append(vid_name.decode())
        t_start_lst.append(segment_predict[i, 0])
        t_end_lst.append(segment_predict[i, 1])
        score_lst.append(segment_predict[i, 2])
        label_lst.append(segment_predict[i, 3])
    prediction = pd.DataFrame(
        {
            "video-id": video_lst,
            "t-start": t_start_lst,
            "t-end": t_end_lst,
            "label": label_lst,
            "score": score_lst,
        }
    )
    return prediction