def _get_64_anchors_new(self, gtbox):
pos = {}
neg = {}
zero = {}
for a in range(17):
for b in range(17):
for c in range(5):
anchor = [self.grid_len//2+self.grid_len*a, self.grid_len//2+self.grid_len*b, self.anchor_shape[c][0], self.anchor_shape[c][1]]
anchor = xywh_to_x1y1x2y2(anchor)
anchor = clip_anchor(anchor,cfg.detection_size)
iou = self._IOU(anchor, gtbox)
# if iou>0.3:
# print(x1y1x2y2_to_xywh(anchor), gtbox, iou)
if iou >= cfg.pos_iou_thresh:
pos[(a,b,c)] = (iou, anchor)
elif iou <= cfg.neg_iou_thresh and iou > 0:
neg[(a,b,c)] = (iou, anchor)
else :
zero[(a,b,c)] = (iou, anchor)
# print(len(pos), len(neg), len(zero))
pos = sorted(pos.items(), key=lambda tup: tup[1][0], reverse=True)
pos = [(*tup[0], tup[1][1]) for tup in pos[:16]]
num = math.ceil((64-len(pos))/2)
neg = sorted(neg.items(), key=lambda tup: tup[1][0], reverse=True)
neg = [(*tup[0], tup[1][1]) for tup in neg[:num]]
zero = list(zero.items())
np.random.shuffle(zero)
neg += [(*tup[0], tup[1][1]) for tup in zero[:64-len(pos)-len(neg)]]
return pos, neg
def _get_label(self, gtbox):
clabel = np.zeros([5, 17, 17]) - 100
rlabel = np.zeros([20, 17, 17], dtype = np.float32)
if self.phase in ['train','validation']:
pos, neg = self._get_64_anchors_new(gtbox)
assert len(pos)+len(neg)==64
for a,b,c,anchor_x1y1x2y2 in pos:
anchor = x1y1x2y2_to_xywh(anchor_x1y1x2y2)
clabel[c,a,b] = 1
channel0 = (gtbox[0] - anchor[0])/anchor[2]
channel1 = (gtbox[1] - anchor[1])/anchor[3]
channel2 = math.log(gtbox[2]/anchor[2])
channel3 = math.log(gtbox[3]/anchor[3])
rlabel[c*4:c*4+4,a,b] = [channel0, channel1, channel2, channel3]
for a,b,c,anchor in neg:
clabel[c,a,b] = 0
else :
for a in range(17):
for b in range(17):
for c in range(5):
anchor = [self.grid_len//2+self.grid_len*a, self.grid_len//2+self.grid_len*b, self.anchor_shape[c][0], self.anchor_shape[c][1]]
anchor = xywh_to_x1y1x2y2(anchor)
anchor = clip_anchor(anchor,cfg.detection_size)
iou = self._IOU(anchor, gtbox)
anchor = x1y1x2y2_to_xywh(anchor)
if iou >= cfg.pos_iou_thresh:
clabel[c,a,b] = 1
channel0 = (gtbox[0] - anchor[0])/anchor[2]
channel1 = (gtbox[1] - anchor[1])/anchor[3]
channel2 = math.log(gtbox[2]/anchor[2])
channel3 = math.log(gtbox[3]/anchor[3])
rlabel[c*4:c*4+4,a,b] = [channel0, channel1, channel2, channel3]
elif iou <= cfg.neg_iou_thresh:
clabel[c,a,b] = 0
return torch.Tensor(clabel).long(), torch.Tensor(rlabel).float()
def _get_64_anchors_multiflags(self,idx,gtboxList):
pos = {}
neg = {}
zero = {}
brother = {}
for a in range(17):
for b in range(17):
for c in range(5):
anchor = [self.grid_len//2+self.grid_len*a, self.grid_len//2+self.grid_len*b, self.anchor_shape[c][0], self.anchor_shape[c][1]]
anchor_x1y1x2y2 = xywh_to_x1y1x2y2(anchor)
anchor_x1y1x2y2 = clip_anchor(anchor_x1y1x2y2,cfg.detection_size)
iouList = list(map(lambda gtbox_xywh: self._IOU(anchor_x1y1x2y2, gtbox_xywh) , gtboxList))
if iouList[idx]>cfg.pos_iou_thresh:
pos[(a,b,c)] = (iou,anchor_x1y1x2y2)
else:
iou = max(iouList[:idx]+iouList[idx+1:])
if iou > cfg.pos_iou_thresh:
brother[(a,b,c)] = (iou,anchor_x1y1x2y2)
iou = max(iou, iouList[idx])
if iou <= cfg.neg_iou_thresh and iou > 0:
neg[(a,b,c)] = (iou, anchor_x1y1x2y2)
elif iou == 0:
zero[(a,b,c)] = (iou, anchor_x1y1x2y2)
pos = sorted(pos.items(), key=lambda tup: tup[1][0], reverse=True)
pos = [(*tup[0], tup[1][1]) for tup in pos[:16]] # list of (a,b,c,anchor_x1y1x2y2)
brother = sorted(brother.items(), key=lambda tup: tup[1][0], reverse=True)
brother = [(*tup[0], tup[1][1]) for tup in brother[:16]]
num = math.ceil((64-len(pos)-len(brother))/2)
zero = list(zero.items())
np.random.shuffle(zero)
zero = [(*tup[0], tup[1][1]) for tup in zero[:num]]
neg = sorted(neg.items(), key=lambda tup: tup[1][0], reverse=True)
neg = [(*tup[0], tup[1][1]) for tup in neg[:64-len(pos)-len(brother)-len(zero)]]
neg += brother + zero
return pos, neg
def outputConvertor(coutput, routput, anchor_shape):
coutput = coutput.detach().cpu().numpy()
routput = routput.detach().cpu().numpy().reshape(5, 4, 17, 17)
bboxList = []
maxProb = 0
for a in range(17):
for b in range(17):
for c in range(5):
anchor = [
cfg.grid_len // 2 + cfg.grid_len * a,
cfg.grid_len // 2 + cfg.grid_len * b, anchor_shape[c][0],
anchor_shape[c][1]
]
anchor_x1y1x2y2 = xywh_to_x1y1x2y2(anchor)
anchor_x1y1x2y2 = clip_anchor(anchor_x1y1x2y2,
cfg.detection_size)
anchor = x1y1x2y2_to_xywh(anchor_x1y1x2y2)
delta = coutput[c, 0, a, b] - coutput[c, 1, a, b]
if delta > 10:
prob = 0
elif delta < -10:
prob = 1
else:
prob = 1 / (1 + math.exp(delta))
maxProb = max(maxProb, prob)
if prob >= 0.5:
# if coutput[c,1,a,b]>0.2:
bbox = [0, 0, 0, 0]
channel0, channel1, channel2, channel3 = routput[c, :, a,
b]
bbox[0] = channel0 * anchor[2] + anchor[0]
bbox[1] = channel1 * anchor[3] + anchor[1]
bbox[2] = math.exp(channel2) * anchor[2]
bbox[3] = math.exp(channel3) * anchor[3]
bbox = xywh_to_x1y1x2y2(bbox)
try:
bbox = np.array(bbox, dtype=np.int32)
except:
print(channel0, channel1, channel2, channel3)
print(bbox)
continue
# raise OverflowError
bboxList.append((bbox, prob))
bboxList.sort(key=lambda tup: tup[1], reverse=True)
return list(map(lambda tup: tup[0], bboxList))[:5], maxProb
def _IOU(self, a, b):
# a = xywh_to_x1y1x2y2(a)
b = xywh_to_x1y1x2y2(b)
sa = (a[2] - a[0]) * (a[3] - a[1])
sb = (b[2] - b[0]) * (b[3] - b[1])
w = max(0, min(a[2], b[2]) - max(a[0], b[0]))
h = max(0, min(a[3], b[3]) - max(a[1], b[1]))
area = w * h
return area / (sa + sb - area)