def evaluate_review(env, image, nms_params, review):
model = env.best.model.to(env.device)
encoder = env.encoder.to(env.device)
scale = env.args.scale
detections = evaluate.evaluate_image(model,
image,
encoder,
device=env.device,
nms_params=nms_params).detections
if detections._size == 0:
return make_detections(env, [])
review = tensors_to(review, device=env.device)
ious = box.iou_matrix(detections.bbox, review.bbox * scale)
ious[ious < nms_params.nms].fill_(-1)
scores = ious.mul(detections.confidence.unsqueeze(1))
review_inds = scores.max(0).values.argsort(descending=True)
detections = table_list(detections)
for i in review_inds.tolist():
score, ind = scores[:, i].max(0)
if score > 0:
detections[ind].match = i
scores[ind].fill_(0)
return make_detections(env, detections)
def encode(target, anchor_boxes, params):
n = anchor_boxes.size(0)
m = target.bbox.size(0)
if m == 0:
return struct(location=target.bbox.new_zeros(n, 4),
classification=target.bbox.new_zeros(n,
dtype=torch.long))
ious = box.iou_matrix(box.point_form(anchor_boxes), target.bbox)
if params.top_anchors > 0:
top_ious, inds = ious.topk(params.top_anchors, dim=0)
ious = ious.scatter(0, inds, top_ious * 2)
max_ious, max_ids = ious.max(1)
class_target = encode_classes(target.label,
max_ious,
max_ids,
match_thresholds=params.match_thresholds)
location = target.bbox[max_ids]
if params.location_loss == "l1":
location = encode_boxes(location, anchor_boxes)
return struct(location=location, classification=class_target)
def match_positives(detections, target):
assert detections.label.dim() == 1 and target.label.dim() == 1
n, m = detections._size, target._size
if m == 0 or n == 0:
return const(torch.FloatTensor(n).zero_())
ious = box.iou_matrix(detections.bbox, target.bbox)
return lambda threshold: _match_positives(
detections.label, target.label, ious, threshold=threshold)
def match_targets(target1, target2, threshold=0.5):
ious = box.iou_matrix(target1.bbox, target2.bbox)
matches = []
if ious.size(0) == 0 or ious.size(1) == 0:
return []
value, loc = take_max_2d(ious)
while value > threshold:
matches.append(struct(match=loc, iou=value))
value, loc = take_max_2d(ious)
return matches
def match_boxes(prediction, target, threshold=0.5, eps=1e-7):
n = prediction.label.size(0)
matches = []
ious = box.iou_matrix(prediction.bbox, target.bbox)
for i, p in enumerate(prediction._sequence()):
match = None
if ious.size(1) > 0:
iou, j = max_1d(ious[i])
label = target.label[j]
matches_box = iou > threshold
if matches_box:
ious[:,
j] = 0 # mark target overlaps to 0 so they won't be selected twice
if p.label == label:
match = (j, iou)
matches.append(p._extend(match=match))
return matches