File size: 4,104 Bytes
5f028d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import cv2
from tqdm import tqdm
import numpy as np
import torch

from ultralytics import YOLO


if torch.cuda.is_available():
    autocast = torch.cuda.amp.autocast
else:
    class autocast:
        def __init__(self, enabled=True):
            pass
        def __enter__(self):
            pass
        def __exit__(self, *args):
            pass


def detect_track(imgfiles, thresh=0.5):
    
    hand_det_model = YOLO('./weights/external/detector.pt')

    # Run
    boxes_ = []
    tracks = {}
    for t, imgpath in enumerate(tqdm(imgfiles)):
        img_cv2 = cv2.imread(imgpath)

        ### --- Detection ---
        with torch.no_grad():
            with autocast():
                results = hand_det_model.track(img_cv2, conf=thresh, persist=True, verbose=False)
                
                boxes = results[0].boxes.xyxy.cpu().numpy()
                confs = results[0].boxes.conf.cpu().numpy()
                handedness = results[0].boxes.cls.cpu().numpy()
                if not results[0].boxes.id is None:
                    track_id = results[0].boxes.id.cpu().numpy()
                else:
                    track_id = [-1] * len(boxes)

                boxes = np.hstack([boxes, confs[:, None]])
                find_right = False
                find_left = False
                for idx, box in enumerate(boxes):
                    if track_id[idx] == -1:
                        if handedness[[idx]] > 0:
                            id = int(10000)
                        else:
                            id = int(5000)
                    else:
                        id = track_id[idx]
                    subj = dict()
                    subj['frame'] = t 
                    subj['det'] = True
                    subj['det_box'] = boxes[[idx]]
                    subj['det_handedness'] = handedness[[idx]]
                    
                    
                    if (not find_right and handedness[[idx]] > 0) or (not find_left and handedness[[idx]]==0):
                        if id in tracks:
                            tracks[id].append(subj)
                        else:
                            tracks[id] = [subj]

                        if handedness[[idx]] > 0:
                            find_right = True
                        elif handedness[[idx]] == 0:
                            find_left = True
    tracks = np.array(tracks, dtype=object)
    boxes_ = np.array(boxes_, dtype=object)

    return boxes_, tracks


def parse_chunks(frame, boxes, min_len=16):
    """ If a track disappear in the middle, 
     we separate it to different segments to estimate the HPS independently. 
     If a segment is less than 16 frames, we get rid of it for now. 
     """
    frame_chunks = []
    boxes_chunks = []
    step = frame[1:] - frame[:-1]
    step = np.concatenate([[0], step])
    breaks = np.where(step != 1)[0]

    start = 0
    for bk in breaks:
        f_chunk = frame[start:bk]
        b_chunk = boxes[start:bk]
        start = bk
        if len(f_chunk)>=min_len:
            frame_chunks.append(f_chunk)
            boxes_chunks.append(b_chunk)

        if bk==breaks[-1]:  # last chunk
            f_chunk = frame[bk:]
            b_chunk = boxes[bk:]
            if len(f_chunk)>=min_len:
                frame_chunks.append(f_chunk)
                boxes_chunks.append(b_chunk)

    return frame_chunks, boxes_chunks

def parse_chunks_hand_frame(frame):
    """ If a track disappear in the middle, 
     we separate it to different segments to estimate the HPS independently. 
     If a segment is less than 16 frames, we get rid of it for now. 
     """
    frame_chunks = []
    step = frame[1:] - frame[:-1]
    step = np.concatenate([[0], step])
    breaks = np.where(step != 1)[0]

    start = 0
    for bk in breaks:
        f_chunk = frame[start:bk]
        start = bk
        if len(f_chunk) > 0:
            frame_chunks.append(f_chunk)

        if bk==breaks[-1]:  # last chunk
            f_chunk = frame[bk:]
            if len(f_chunk) > 0:
                frame_chunks.append(f_chunk)

    return frame_chunks