def frame_callback(vis, frame_idx):
print("Frame idx", frame_idx)
image = cv2.imread(seq_info["image_filenames"][frame_idx],
cv2.IMREAD_COLOR)
vis.set_image(image.copy())
if seq_info["detections"] is not None:
detections = deep_sort_app.create_detections(
seq_info["detections"], frame_idx)
vis.draw_detections(detections)
mask = results[:, 0].astype(np.int) == frame_idx
track_ids = results[mask, 1].astype(np.int)
boxes = results[mask, 2:6]
vis.draw_groundtruth(track_ids, boxes)
if show_false_alarms:
groundtruth = seq_info["groundtruth"]
mask = groundtruth[:, 0].astype(np.int) == frame_idx
gt_boxes = groundtruth[mask, 2:6]
for box in boxes:
# NOTE(nwojke): This is not strictly correct, because we don't
# solve the assignment problem here.
min_iou_overlap = 0.5
if iou(box, gt_boxes).max() < min_iou_overlap:
vis.viewer.color = 0, 0, 255
vis.viewer.thickness = 4
vis.viewer.rectangle(*box.astype(np.int))
def frame_callback(vis, frame_idx):
print("Frame idx", frame_idx)
image = cv2.imread(
seq_info["image_filenames"][frame_idx], cv2.IMREAD_COLOR)
vis.set_image(image.copy())
if seq_info["detections"] is not None:
detections = deep_sort_app.create_detections(
seq_info["detections"], frame_idx)
vis.draw_detections(detections)
mask = results[:, 0].astype(np.int) == frame_idx
track_ids = results[mask, 1].astype(np.int)
boxes = results[mask, 2:6]
vis.draw_groundtruth(track_ids, boxes)
if show_false_alarms:
groundtruth = seq_info["groundtruth"]
mask = groundtruth[:, 0].astype(np.int) == frame_idx
gt_boxes = groundtruth[mask, 2:6]
for box in boxes:
# NOTE(nwojke): This is not strictly correct, because we don't
# solve the assignment problem here.
min_iou_overlap = 0.5
if iou(box, gt_boxes).max() < min_iou_overlap:
vis.viewer.color = 0, 0, 255
vis.viewer.thickness = 4
vis.viewer.rectangle(*box.astype(np.int))
def frame_callback(vis, frame_idx):
#프레임별로 처리
print("Frame idx", frame_idx)
image = cv2.imread(
seq_info["image_filenames"][frame_idx], cv2.IMREAD_COLOR)
vis.set_image(image.copy())
if seq_info["detections"] is not None:
detections = deep_sort_app.create_detections(
seq_info["detections"], frame_idx)
vis.draw_detections(detections)
mask = results[:, 0].astype(np.int) == frame_idx
track_ids = results[mask, 1].astype(np.int) #해당 frame_id인 mask값들 중 [1]들인 id값 추출
boxes = results[mask, 2:6]
vis.draw_groundtruth(track_ids, boxes)
# 발위치 10개 중 y값만 빼기
h_file = os.path.dirname(result_file) # result/text/
with open(h_file + '/ID_h.txt', 'r') as f_hi:
line_splits = [int(l.split(',')[1]) for l in f_hi.read().splitlines()[1:]]
i = 0
#print(line_splits)
if show_false_alarms:
groundtruth = seq_info["groundtruth"]
mask = groundtruth[:, 0].astype(np.int) == frame_idx
gt_boxes = groundtruth[mask, 2:6]
for box in boxes:
# NOTE(nwojke): This is not strictly correct, because we don't
# solve the assignment problem here.
min_iou_overlap = 0.5
if iou(box, gt_boxes).max() < min_iou_overlap:
vis.viewer.color = 0, 0, 255
vis.viewer.thickness = 4
vis.viewer.rectangle(*box.astype(np.int))
if IDnum != 0: # Tracking 하는 ID만 보여주고 발 표시 !!!!!!!
vis.viewer.circle(
box[0] + box[2] / 2, box[1] + box[3], 3)
def process_next_frame(self, frame):
"""
Track objects from specified detections
:param frame: frame data + list of detections, map-like object with mandatory keys: image, detections
:return: detections populated with object ids
"""
if len(frame['detections']['rois']) == 0:
return frame
frame['detections']['rois'][:, 2] -= frame['detections']['rois'][:, 0]
frame['detections']['rois'][:, 3] -= frame['detections']['rois'][:, 1]
frame['detections']['features'] = self.__feature_extractor(
frame['image'], frame['detections']['rois'])
self.__tracker.predict()
self.__tracker.update([
Detection(frame['detections']['rois'][idx],
frame['detections']['scores'][idx],
frame['detections']['features'][idx])
for idx, d in enumerate(frame['detections']['rois'])
])
tracked_bbox = [
track.to_tlwh() for track in self.__tracker.tracks
if track.state == TrackState.Confirmed
]
for idx_detection, detection in enumerate(frame['detections']['rois']):
for idx_track, track in enumerate(tracked_bbox):
if iou(
tracked_bbox[idx_track],
np.array(frame['detections']['rois'][idx_detection],
dtype=np.float).reshape(1, 4))[0] >= 0.7:
frame['detections']['ids'][
idx_detection] = self.__tracker.tracks[
idx_track].track_id
return frame
def draw_trackers(self, tracks, gts):
#TODO make another option where you show any track which overlaps with the given object
# this will require the groundtruth
SHOW_OVERLAPPED = True
if self.vis_method == "one-gt":
if tracks == []: # there's issues with zero lists
return
# for convenience we'll use self.gt_to_vis and self.tracks_to_vis
# perhaps this isn't the way to do it. There will need to be a single gt index and a set of track indices
# the logic is going to be pretty funky here
gt_ids = gts[0]
gt_boxs = gts[1]
assert len(gt_ids) == len(gt_boxs)
if (len(gt_ids) > 0 and self.gt_to_vis is None) or\
(self.gt_to_vis is not None and self.gt_to_vis not in gt_ids and len(gt_ids) > 0): # there are two cases to change the gt_ids, either it is unset or the one we were tracking is no longer present
# the groundtruth should be in the form (List(ids), List(boxes))
self.gt_to_vis = int(random.choice(gt_ids))
self.tracks_to_vis = []
if self.gt_to_vis is not None and len(gt_ids) > 0:
# here we want to add any tracks that overlap at all and remove a
# TODO run through all of the tracks and see if they overlap with the selceted gt
# TODO determine which of the bboxs coresponds to the index that's being visualized
# find the index of self.gt_to_vis in groundtruths[0]
gt_box = gt_boxs[gt_ids.index(
self.gt_to_vis
)] # find the box which coresponds to the index we are visualizing
track_boxes = np.asarray([t.to_tlwh() for t in tracks])
track_indices = np.asarray([t.track_id for t in tracks])
overlaps = iou(gt_box, track_boxes)
# these should be tracks which overlap with the groundtruth track we've picked
# TODO determine why additional tracks are being added
new_tracks_to_vis = track_indices[np.nonzero(overlaps)[
0]] # for some reason this return a tuple of arrays
def union(a, b):
""" return the union of two lists """
return list(set(a) | set(b))
self.tracks_to_vis = union(self.tracks_to_vis,
new_tracks_to_vis)
tracks_ = [
t for t in tracks if t.track_id in self.tracks_to_vis
] # I don't want to change `tracks` as it was passed by reference
for track in tracks_:
if not track.is_confirmed(): # or track.time_since_update > 0:
continue
self.viewer.color = create_unique_color_uchar(track.track_id)
if track.time_since_update > 0:
self.viewer.thickness = 2
else:
self.viewer.thickness = 5
self.viewer.rectangle(*track.to_tlwh().astype(np.int),
label=str(track.track_id))
elif self.vis_method == "one-track": # I believe these tracks are sorted w.r.t. to seniority, so this should handle it niavely
# check if the one we want to visualize
confirmed_ids = [
track.track_id for track in tracks if track.is_confirmed()
]
if self.index_to_vis not in confirmed_ids and len(
confirmed_ids) > 0: # the track must have died
self.index_to_vis = random.choice(confirmed_ids)
tracks = [t for t in tracks if t.track_id == self.index_to_vis
] # this is the cleanest way I found to get the item
if len(tracks) == 0:
return
track = tracks[0]
self.viewer.color = create_unique_color_uchar(track.track_id)
if track.time_since_update > 0:
self.viewer.thickness = 2
else:
self.viewer.thickness = 5
self.viewer.rectangle(*track.to_tlwh().astype(np.int),
label=str(track.track_id))
elif self.vis_method == "show-all":
for track in tracks:
#HACK
#if not track.is_confirmed():# or track.time_since_update > 0:
# continue
self.viewer.color = create_unique_color_uchar(track.track_id)
if track.time_since_update > 0:
self.viewer.thickness = 2
else:
self.viewer.thickness = 5
self.viewer.rectangle(*track.to_tlwh().astype(np.int),
label=str(track.track_id))
#if track.time_since_update > 0:
# self.viewer.color = (255, 255, 255)
# self.viewer.thickness = 2
# self.viewer.rectangle(
# *track.to_tlwh().astype(np.int), label=str(track.track_id))
#self.viewer.gaussian(track.mean[:2], track.covariance[:2, :2],
# label="%d" % track.track_id)
else:
raise ValueError(
"self.vis_method should be `show all`, `one-track`, or `one-gt` but insted was {}"
.format(self.vis_method))
gt_ids, gt_boxes = zip(
*[(g['local_id'], _tlbr_to_tlwh(g['box']))
for g in gt[frame_idx]]) if len(gt[frame_idx]) > 0 else ([],
[])
det_ids, det_boxes = zip(
*[(d['local_id'], _tlbr_to_tlwh(d['location']))
for d in context[frame_idx]['context']]) if len(
context[frame_idx]['context']) > 0 else ([], [])
candidates = np.asarray(det_boxes)
gt_boxes = np.asarray(gt_boxes)
costs = []
for gt_box in gt_boxes:
if len(candidates) == 0:
costs.append([])
else:
cost = 1. - iou(gt_box, candidates)
cost = list(
map(lambda x: np.nan
if x >= args.iou_cost_th else x, cost))
costs.append(cost)
acc.update(gt_ids, det_ids, costs)
# 集計
df_sub = mh.compute(acc, metrics=metrics, name=video_name)
if df is None:
df = df_sub
else:
df = pd.concat([df, df_sub])
# motmetric出力
pkl_filepath = os.path.join(args.save_root, video_name, 'motacc.pkl')
with open(pkl_filepath, 'wb') as f:
pickle.dump(acc, f)