def openpose_detections(self,
camera,
start_frame=1,
stop_frame=float('Inf')):
detections = h5py.File(self.path +
'/detections/openpose/camera%d.mat' %
camera)['detections']
i0 = bisect(RowView(detections, 1), start_frame) - 1
while detections[1, i0] == start_frame and i0 >= 0:
i0 -= 1
i0 += 1
for ind in range(i0, detections.shape[1]):
det = detections[:, ind]
camera = int(det[0])
frame = int(det[1])
left, top, width, height = pose2bb(det[2:])
left, top, right, bottom = map(
lambda x: int(self.scale * x),
[left, top, left + width, top + height])
confidence = det[-1]
if frame > stop_frame:
break
det = Detection(frame, left, top, right, bottom, confidence, ind)
det.scene_name = camera
yield det
def detections(self, start_frame=1, stop_frame=np.inf):
if self._detections is None:
detections = defaultdict(list)
part, seq = self.name.split('__')
seq = os.path.join(self.dataset.base_path,
self.dataset.detections_dir, part, seq + '.txt')
did = 0
for l in open(seq, "r").readlines():
frow = tuple(map(float, l.split(',')))
row = tuple(map(int, frow))
if row[7] in self.dataset.class_indexes:
det = Detection(row[0], row[2], row[3], row[2] + row[4],
row[3] + row[5], frow[6], did)
if self.dataset.scale != 1.0:
det = det.scale(self.dataset.scale)
if not self.should_ignore(det):
det.cls = row[7]
det.scene_name = self.name
detections[det.frame].append(det)
did += 1
for dets in detections.values():
dets[:] = nms(dets)
self._detections = detections
self._last_frame = max(detections.keys())
stop_frame = min(stop_frame, self._last_frame)
for f in range(start_frame, stop_frame + 1):
for det in self._detections[f]:
yield det
def ground_truth_detections(self, camera):
gt = io.loadmat(self.path + '/ground_truth/trainval.mat')['trainData']
gt = gt[gt[:, 2].argsort()]
for cam, track_id, frame, left, top, width, height, worldX, worldY, feetX, feetyY in gt:
if cam != camera:
continue
det = Detection(frame, left, top, left + width, top + height, None,
int(track_id))
if self.scale != 1.0:
det = det.scale(self.scale)
yield det
def ground_truth(self, scene, classes=None):
if classes is None:
class_indexes = self.class_indexes
else:
class_indexes = set([self.class_names.index(c) for c in classes])
frames = defaultdict(list)
fn = self._path(scene, 'annotations') + '.txt'
for row in np.loadtxt(fn, delimiter=',', dtype=int):
if row[7] in class_indexes:
det = Detection(row[0], row[2], row[3], row[2] + row[4],
row[3] + row[5], None, row[1])
det.cls = row[7]
frames[det.frame].append(det)
return frames
def ground_truth(self, scene):
frames = defaultdict(list)
for row in np.loadtxt(self._path(scene, "gt/gt.txt"), delimiter=','):
if row[0] < 1 or row[6] == 0 or row[7] != 1:
continue
det = Detection(int(row[0]), row[2], row[3], row[2] + row[4],
row[3] + row[5], None, int(row[1]))
frames[det.frame].append(det)
return frames
def detections(self, start_frame=1, stop_frame=np.inf):
if self._detections is None:
detections = defaultdict(list)
did = 0
for row in np.loadtxt(self.dataset._path(self.name, "det/det.txt"),
delimiter=','):
if row[0] < 1 or row[6] == 0:
print(row[0], row[6], row[7])
continue
det = Detection(int(row[0]), row[2], row[3], row[2] + row[4],
row[3] + row[5], row[6], did)
det.scene_name = self.name
detections[det.frame].append(det)
did += 1
for dets in detections.values():
dets[:] = nms(dets)
self._detections = detections
self._last_frame = max(detections.keys())
stop_frame = min(stop_frame, self._last_frame)
for f in range(start_frame, stop_frame + 1):
for det in self._detections[f]:
yield det
def test_detetcion(self):
d1 = Detection(0, 10, 20, 100, 200, 0.9, 7)
assert d1.cx == 55
assert d1.cy == 110
assert d1.area == 16200
d2 = Detection(2, 20, 30, 110, 220, 0.9, 7)
assert d1.iou(d2) == d2.iou(d1) == slow_iou(d1, d2)
assert d1.iou(d1) == d2.iou(d2) == slow_iou(d1, d1) == slow_iou(
d1, d1) == 1.0
assert d1.ioa(d2) == 0.8395061728395061
assert d2.ioa(d1) == 0.7953216374269005
assert d1.distance_to(d2) == d2.distance_to(d1) == 18.027756377319946
assert d1.interpolate(d2, 1) == (1, 15, 25, 105, 210, None, None)
assert d1.predict(2, 3, 4) == (2, 16, 28, 106, 208, None, None)
assert d1.scale(2) == (0, 20, 40, 200, 400, 0.9, 7)
assert d1.move(2, 3) == (0, 12, 23, 102, 203, 0.9, 7)
mask = np.zeros((500, 500))
d1.update_mask(mask)
d2.update_mask(mask)
assert mask.sum() == 5023500
assert mask[0, 0] == mask[5, 50] == mask[50, 5] == 0
assert mask[20, 10] == 255
assert mask[199, 99] == 255
for _ in range(1000):
d1 = random_detection()
d2 = random_detection()
assert d1.iou(d2) == d2.iou(d1) == slow_iou(d1, d2)
assert d1.covers(d1.cx, d1.cy)
assert d2.covers(d2.cx, d2.cy)
mask = np.zeros((500, 500))
d1.update_mask(mask)
d2.update_mask(mask)
assert mask[int(d1.cy), int(d1.cx)] == 255
assert mask[int(d2.cy), int(d2.cx)] == 255
def random_detection():
l = randrange(100)
t = randrange(100)
w = randrange(1, 100)
h = randrange(1, 100)
return Detection(0, l, t, l + w, t + h, 0.9, 7)