def match_targets_to_proposals(self, proposal, target):
match_quality_matrix = boxlist_iou(target, proposal)
matched_idxs = self.proposal_matcher(match_quality_matrix)
# Fast RCNN only need "labels" field for selecting the targets
target = target.copy_with_fields("labels")
# get the targets corresponding GT for each proposal
# NB: need to clamp the indices because we can have a single
# GT in the image, and matched_idxs can be -2, which goes
# out of bounds
matched_targets = target[matched_idxs.clamp(min=0)]
matched_targets.add_field("matched_idxs", matched_idxs)
return matched_targets
def match_targets_to_anchors(self, anchor, target, copied_fields=[]):
match_quality_matrix = boxlist_iou(target, anchor)
matched_idxs = self.proposal_matcher(match_quality_matrix)
# RPN doesn't need any fields from target
# for creating the labels, so clear them all
target = target.copy_with_fields(copied_fields)
# get the targets corresponding GT for each anchor
# NB: need to clamp the indices because we can have a single
# GT in the image, and matched_idxs can be -2, which goes
# out of bounds
matched_targets = target[matched_idxs.clamp(min=0)]
matched_targets.add_field("matched_idxs", matched_idxs)
return matched_targets
def evaluate_box_proposals(predictions,
dataset,
thresholds=None,
area="all",
limit=None):
"""Evaluate detection proposal recall metrics. This function is a much
faster alternative to the official COCO API recall evaluation code. However,
it produces slightly different results.
"""
# Record max overlap value for each gt box
# Return vector of overlap values
areas = {
"all": 0,
"small": 1,
"medium": 2,
"large": 3,
"96-128": 4,
"128-256": 5,
"256-512": 6,
"512-inf": 7
}
area_ranges = [
[0**2, 1e5**2], # all
[0**2, 32**2], # small
[32**2, 96**2], # medium
[96**2, 1e5**2], # large
[96**2, 128**2], # 96-128
[128**2, 256**2], # 128-256
[256**2, 512**2], # 256-512
[512**2, 1e5**2]
] # 512-inf
assert area in areas, "Unknown area range: {}".format(area)
area_range = area_ranges[areas[area]]
gt_overlaps = []
num_pos = 0
for image_id, prediction in enumerate(predictions):
original_id = dataset.id_to_img_map[image_id]
img_info = dataset.get_img_info(image_id)
image_width = img_info["width"]
image_height = img_info["height"]
prediction = prediction.resize((image_width, image_height))
# sort predictions in descending order
# TODO maybe remove this and make it explicit in the documentation
inds = prediction.get_field("objectness").sort(descending=True)[1]
prediction = prediction[inds]
ann_ids = dataset.coco.getAnnIds(imgIds=original_id)
anno = dataset.coco.loadAnns(ann_ids)
gt_boxes = [obj["bbox"] for obj in anno if obj["iscrowd"] == 0]
gt_boxes = torch.as_tensor(gt_boxes).reshape(
-1, 4) # guard against no boxes
gt_boxes = BoxList(gt_boxes, (image_width, image_height),
mode="xywh").convert("xyxy")
gt_areas = torch.as_tensor(
[obj["area"] for obj in anno if obj["iscrowd"] == 0])
if len(gt_boxes) == 0:
continue
valid_gt_inds = (gt_areas >= area_range[0]) & (gt_areas <=
area_range[1])
gt_boxes = gt_boxes[valid_gt_inds]
num_pos += len(gt_boxes)
if len(gt_boxes) == 0:
continue
if len(prediction) == 0:
continue
if limit is not None and len(prediction) > limit:
prediction = prediction[:limit]
overlaps = boxlist_iou(prediction, gt_boxes)
_gt_overlaps = torch.zeros(len(gt_boxes))
for j in range(min(len(prediction), len(gt_boxes))):
# find which proposal box maximally covers each gt box
# and get the iou amount of coverage for each gt box
max_overlaps, argmax_overlaps = overlaps.max(dim=0)
# find which gt box is 'best' covered (i.e. 'best' = most iou)
gt_ovr, gt_ind = max_overlaps.max(dim=0)
assert gt_ovr >= 0
# find the proposal box that covers the best covered gt box
box_ind = argmax_overlaps[gt_ind]
# record the iou coverage of this gt box
_gt_overlaps[j] = overlaps[box_ind, gt_ind]
assert _gt_overlaps[j] == gt_ovr
# mark the proposal box and the gt box as used
overlaps[box_ind, :] = -1
overlaps[:, gt_ind] = -1
# append recorded iou coverage level
gt_overlaps.append(_gt_overlaps)
gt_overlaps = torch.cat(gt_overlaps, dim=0)
gt_overlaps, _ = torch.sort(gt_overlaps)
if thresholds is None:
step = 0.05
thresholds = torch.arange(0.5, 0.95 + 1e-5, step, dtype=torch.float32)
recalls = torch.zeros_like(thresholds)
# compute recall for each iou threshold
for i, t in enumerate(thresholds):
recalls[i] = (gt_overlaps >= t).float().sum() / float(num_pos)
# ar = 2 * np.trapz(recalls, thresholds)
ar = recalls.mean()
return {
"ar": ar,
"recalls": recalls,
"thresholds": thresholds,
"gt_overlaps": gt_overlaps,
"num_pos": num_pos
}
def prepare_targets(self, targets, anchors):
cls_labels = []
reg_targets = []
for im_i in range(len(targets)):
targets_per_im = targets[im_i]
assert targets_per_im.mode == "xyxy"
bboxes_per_im = targets_per_im.bbox
labels_per_im = targets_per_im.get_field("labels")
anchors_per_im = cat_boxlist(anchors[im_i])
num_gt = bboxes_per_im.shape[0]
# ATSS sampling----------------------------
num_anchors_per_level = [len(anchors_per_level.bbox) for anchors_per_level in anchors[im_i]]
ious = boxlist_iou(anchors_per_im, targets_per_im)
gt_cx = (bboxes_per_im[:, 2] + bboxes_per_im[:, 0]) / 2.0
gt_cy = (bboxes_per_im[:, 3] + bboxes_per_im[:, 1]) / 2.0
gt_points = torch.stack((gt_cx, gt_cy), dim=1)
anchors_cx_per_im = (anchors_per_im.bbox[:, 2] + anchors_per_im.bbox[:, 0]) / 2.0
anchors_cy_per_im = (anchors_per_im.bbox[:, 3] + anchors_per_im.bbox[:, 1]) / 2.0
anchor_points = torch.stack((anchors_cx_per_im, anchors_cy_per_im), dim=1)
distances = (anchor_points[:, None, :] - gt_points[None, :, :]).pow(2).sum(-1).sqrt()
# Selecting candidates based on the center distance between anchor box and object
candidate_idxs = []
star_idx = 0
for level, anchors_per_level in enumerate(anchors[im_i]):
end_idx = star_idx + num_anchors_per_level[level]
distances_per_level = distances[star_idx:end_idx, :]
topk = min(self.cfg.top_k, num_anchors_per_level[level])
_, topk_idxs_per_level = distances_per_level.topk(topk, dim=0, largest=False)
candidate_idxs.append(topk_idxs_per_level + star_idx)
star_idx = end_idx
candidate_idxs = torch.cat(candidate_idxs, dim=0)
# Using the sum of mean and standard deviation as the IoU threshold to select final positive samples
candidate_ious = ious[candidate_idxs, torch.arange(num_gt)]
iou_mean_per_gt = candidate_ious.mean(0)
iou_std_per_gt = candidate_ious.std(0)
iou_thresh_per_gt = iou_mean_per_gt + iou_std_per_gt
is_pos = candidate_ious >= iou_thresh_per_gt[None, :]
# Limiting the final positive samples’ center to object
anchor_num = anchors_cx_per_im.shape[0]
for ng in range(num_gt):
candidate_idxs[:, ng] += ng * anchor_num
e_anchors_cx = anchors_cx_per_im.view(1, -1).expand(num_gt, anchor_num).contiguous().view(-1)
e_anchors_cy = anchors_cy_per_im.view(1, -1).expand(num_gt, anchor_num).contiguous().view(-1)
candidate_idxs = candidate_idxs.view(-1)
l = e_anchors_cx[candidate_idxs].view(-1, num_gt) - bboxes_per_im[:, 0]
t = e_anchors_cy[candidate_idxs].view(-1, num_gt) - bboxes_per_im[:, 1]
r = bboxes_per_im[:, 2] - e_anchors_cx[candidate_idxs].view(-1, num_gt)
b = bboxes_per_im[:, 3] - e_anchors_cy[candidate_idxs].view(-1, num_gt)
is_in_gts = torch.stack([l, t, r, b], dim=1).min(dim=1)[0] > 0.01
is_pos = is_pos & is_in_gts
# if an anchor box is assigned to multiple gts, the one with the highest IoU will be selected.
ious_inf = torch.full_like(ious, -INF).t().contiguous().view(-1)
index = candidate_idxs.view(-1)[is_pos.view(-1)]
ious_inf[index] = ious.t().contiguous().view(-1)[index]
ious_inf = ious_inf.view(num_gt, -1).t()
anchors_to_gt_values, anchors_to_gt_indexs = ious_inf.max(dim=1)
cls_labels_per_im = labels_per_im[anchors_to_gt_indexs]
cls_labels_per_im[anchors_to_gt_values == -INF] = 0
matched_gts = bboxes_per_im[anchors_to_gt_indexs]
reg_targets_per_im = self.box_coder.encode(matched_gts, anchors_per_im.bbox)
cls_labels.append(cls_labels_per_im)
reg_targets.append(reg_targets_per_im)
return cls_labels, reg_targets
def calc_detection_voc_prec_rec(gt_boxlists, pred_boxlists, iou_thresh=0.5):
"""Calculate precision and recall based on evaluation code of PASCAL VOC.
This function calculates precision and recall of
predicted bounding boxes obtained from a dataset which has :math:`N`
images.
The code is based on the evaluation code used in PASCAL VOC Challenge.
"""
n_pos = defaultdict(int)
score = defaultdict(list)
match = defaultdict(list)
for gt_boxlist, pred_boxlist in zip(gt_boxlists, pred_boxlists):
pred_bbox = pred_boxlist.bbox.numpy()
pred_label = pred_boxlist.get_field("labels").numpy()
pred_score = pred_boxlist.get_field("scores").numpy()
gt_bbox = gt_boxlist.bbox.numpy()
gt_label = gt_boxlist.get_field("labels").numpy()
gt_difficult = gt_boxlist.get_field("difficult").numpy()
for l in np.unique(np.concatenate((pred_label, gt_label)).astype(int)):
pred_mask_l = pred_label == l
pred_bbox_l = pred_bbox[pred_mask_l]
pred_score_l = pred_score[pred_mask_l]
# sort by score
order = pred_score_l.argsort()[::-1]
pred_bbox_l = pred_bbox_l[order]
pred_score_l = pred_score_l[order]
gt_mask_l = gt_label == l
gt_bbox_l = gt_bbox[gt_mask_l]
gt_difficult_l = gt_difficult[gt_mask_l]
n_pos[l] += np.logical_not(gt_difficult_l).sum()
score[l].extend(pred_score_l)
if len(pred_bbox_l) == 0:
continue
if len(gt_bbox_l) == 0:
match[l].extend((0,) * pred_bbox_l.shape[0])
continue
# VOC evaluation follows integer typed bounding boxes.
pred_bbox_l = pred_bbox_l.copy()
pred_bbox_l[:, 2:] += 1
gt_bbox_l = gt_bbox_l.copy()
gt_bbox_l[:, 2:] += 1
iou = boxlist_iou(
BoxList(pred_bbox_l, gt_boxlist.size),
BoxList(gt_bbox_l, gt_boxlist.size),
).numpy()
gt_index = iou.argmax(axis=1)
# set -1 if there is no matching ground truth
gt_index[iou.max(axis=1) < iou_thresh] = -1
del iou
selec = np.zeros(gt_bbox_l.shape[0], dtype=bool)
for gt_idx in gt_index:
if gt_idx >= 0:
if gt_difficult_l[gt_idx]:
match[l].append(-1)
else:
if not selec[gt_idx]:
match[l].append(1)
else:
match[l].append(0)
selec[gt_idx] = True
else:
match[l].append(0)
n_fg_class = max(n_pos.keys()) + 1
prec = [None] * n_fg_class
rec = [None] * n_fg_class
for l in n_pos.keys():
score_l = np.array(score[l])
match_l = np.array(match[l], dtype=np.int8)
order = score_l.argsort()[::-1]
match_l = match_l[order]
tp = np.cumsum(match_l == 1)
fp = np.cumsum(match_l == 0)
# If an element of fp + tp is 0,
# the corresponding element of prec[l] is nan.
prec[l] = tp / (fp + tp)
# If n_pos[l] is 0, rec[l] is None.
if n_pos[l] > 0:
rec[l] = tp / n_pos[l]
return prec, rec
def prepare_targets(self, targets, anchors):
cls_labels = []
reg_targets = []
for im_i in range(len(targets)):
targets_per_im = targets[im_i]
assert targets_per_im.mode == "xyxy"
bboxes_per_im = targets_per_im.bbox
labels_per_im = targets_per_im.get_field("labels")
anchors_per_im = cat_boxlist(anchors[im_i])
num_gt = bboxes_per_im.shape[0]
if self.cfg.MODEL.ATSS.POSITIVE_TYPE == 'SSC':
object_sizes_of_interest = [[-1, 64], [64, 128], [128, 256], [256, 512], [512, INF]]
area_per_im = targets_per_im.area()
expanded_object_sizes_of_interest = []
points = []
for l, anchors_per_level in enumerate(anchors[im_i]):
anchors_per_level = anchors_per_level.bbox
anchors_cx_per_level = (anchors_per_level[:, 2] + anchors_per_level[:, 0]) / 2.0
anchors_cy_per_level = (anchors_per_level[:, 3] + anchors_per_level[:, 1]) / 2.0
points_per_level = torch.stack((anchors_cx_per_level, anchors_cy_per_level), dim=1)
points.append(points_per_level)
object_sizes_of_interest_per_level = \
points_per_level.new_tensor(object_sizes_of_interest[l])
expanded_object_sizes_of_interest.append(
object_sizes_of_interest_per_level[None].expand(len(points_per_level), -1)
)
expanded_object_sizes_of_interest = torch.cat(expanded_object_sizes_of_interest, dim=0)
points = torch.cat(points, dim=0)
xs, ys = points[:, 0], points[:, 1]
l = xs[:, None] - bboxes_per_im[:, 0][None]
t = ys[:, None] - bboxes_per_im[:, 1][None]
r = bboxes_per_im[:, 2][None] - xs[:, None]
b = bboxes_per_im[:, 3][None] - ys[:, None]
reg_targets_per_im = torch.stack([l, t, r, b], dim=2)
is_in_boxes = reg_targets_per_im.min(dim=2)[0] > 0.01
max_reg_targets_per_im = reg_targets_per_im.max(dim=2)[0]
is_cared_in_the_level = \
(max_reg_targets_per_im >= expanded_object_sizes_of_interest[:, [0]]) & \
(max_reg_targets_per_im <= expanded_object_sizes_of_interest[:, [1]])
locations_to_gt_area = area_per_im[None].repeat(len(points), 1)
locations_to_gt_area[is_in_boxes == 0] = INF
locations_to_gt_area[is_cared_in_the_level == 0] = INF
locations_to_min_area, locations_to_gt_inds = locations_to_gt_area.min(dim=1)
cls_labels_per_im = labels_per_im[locations_to_gt_inds]
cls_labels_per_im[locations_to_min_area == INF] = 0
matched_gts = bboxes_per_im[locations_to_gt_inds]
elif self.cfg.MODEL.ATSS.POSITIVE_TYPE == 'ATSS':
num_anchors_per_level = [len(anchors_per_level.bbox) for anchors_per_level in anchors[im_i]]
ious = boxlist_iou(anchors_per_im, targets_per_im)
gt_cx = (bboxes_per_im[:, 2] + bboxes_per_im[:, 0]) / 2.0
gt_cy = (bboxes_per_im[:, 3] + bboxes_per_im[:, 1]) / 2.0
gt_points = torch.stack((gt_cx, gt_cy), dim=1)
anchors_cx_per_im = (anchors_per_im.bbox[:, 2] + anchors_per_im.bbox[:, 0]) / 2.0
anchors_cy_per_im = (anchors_per_im.bbox[:, 3] + anchors_per_im.bbox[:, 1]) / 2.0
anchor_points = torch.stack((anchors_cx_per_im, anchors_cy_per_im), dim=1)
distances = (anchor_points[:, None, :] - gt_points[None, :, :]).pow(2).sum(-1).sqrt()
# Selecting candidates based on the center distance between anchor box and object
candidate_idxs = []
star_idx = 0
for level, anchors_per_level in enumerate(anchors[im_i]):
end_idx = star_idx + num_anchors_per_level[level]
distances_per_level = distances[star_idx:end_idx, :]
_, topk_idxs_per_level = distances_per_level.topk(self.cfg.MODEL.ATSS.TOPK, dim=0, largest=False)
candidate_idxs.append(topk_idxs_per_level + star_idx)
star_idx = end_idx
candidate_idxs = torch.cat(candidate_idxs, dim=0)
# Using the sum of mean and standard deviation as the IoU threshold to select final positive samples
candidate_ious = ious[candidate_idxs, torch.arange(num_gt)]
iou_mean_per_gt = candidate_ious.mean(0)
iou_std_per_gt = candidate_ious.std(0)
iou_thresh_per_gt = iou_mean_per_gt + iou_std_per_gt
is_pos = candidate_ious >= iou_thresh_per_gt[None, :]
# Limiting the final positive samples’ center to object
anchor_num = anchors_cx_per_im.shape[0]
for ng in range(num_gt):
candidate_idxs[:, ng] += ng * anchor_num
e_anchors_cx = anchors_cx_per_im.view(1, -1).expand(num_gt, anchor_num).contiguous().view(-1)
e_anchors_cy = anchors_cy_per_im.view(1, -1).expand(num_gt, anchor_num).contiguous().view(-1)
candidate_idxs = candidate_idxs.view(-1)
l = e_anchors_cx[candidate_idxs].view(-1, num_gt) - bboxes_per_im[:, 0]
t = e_anchors_cy[candidate_idxs].view(-1, num_gt) - bboxes_per_im[:, 1]
r = bboxes_per_im[:, 2] - e_anchors_cx[candidate_idxs].view(-1, num_gt)
b = bboxes_per_im[:, 3] - e_anchors_cy[candidate_idxs].view(-1, num_gt)
is_in_gts = torch.stack([l, t, r, b], dim=1).min(dim=1)[0] > 0.01
is_pos = is_pos & is_in_gts
# if an anchor box is assigned to multiple gts, the one with the highest IoU will be selected.
ious_inf = torch.full_like(ious, -INF).t().contiguous().view(-1)
index = candidate_idxs.view(-1)[is_pos.view(-1)]
ious_inf[index] = ious.t().contiguous().view(-1)[index]
ious_inf = ious_inf.view(num_gt, -1).t()
anchors_to_gt_values, anchors_to_gt_indexs = ious_inf.max(dim=1)
cls_labels_per_im = labels_per_im[anchors_to_gt_indexs]
cls_labels_per_im[anchors_to_gt_values == -INF] = 0
matched_gts = bboxes_per_im[anchors_to_gt_indexs]
elif self.cfg.MODEL.ATSS.POSITIVE_TYPE == 'IoU':
match_quality_matrix = boxlist_iou(targets_per_im, anchors_per_im)
matched_idxs = self.matcher(match_quality_matrix)
targets_per_im = targets_per_im.copy_with_fields(['labels'])
matched_targets = targets_per_im[matched_idxs.clamp(min=0)]
cls_labels_per_im = matched_targets.get_field("labels")
cls_labels_per_im = cls_labels_per_im.to(dtype=torch.float32)
# Background (negative examples)
bg_indices = matched_idxs == Matcher.BELOW_LOW_THRESHOLD
cls_labels_per_im[bg_indices] = 0
# discard indices that are between thresholds
inds_to_discard = matched_idxs == Matcher.BETWEEN_THRESHOLDS
cls_labels_per_im[inds_to_discard] = -1
matched_gts = matched_targets.bbox
# Limiting positive samples’ center to object
# in order to filter out poor positives and use the centerness branch
pos_idxs = torch.nonzero(cls_labels_per_im > 0).squeeze(1)
pos_anchors_cx = (anchors_per_im.bbox[pos_idxs, 2] + anchors_per_im.bbox[pos_idxs, 0]) / 2.0
pos_anchors_cy = (anchors_per_im.bbox[pos_idxs, 3] + anchors_per_im.bbox[pos_idxs, 1]) / 2.0
l = pos_anchors_cx - matched_gts[pos_idxs, 0]
t = pos_anchors_cy - matched_gts[pos_idxs, 1]
r = matched_gts[pos_idxs, 2] - pos_anchors_cx
b = matched_gts[pos_idxs, 3] - pos_anchors_cy
is_in_gts = torch.stack([l, t, r, b], dim=1).min(dim=1)[0] > 0.01
cls_labels_per_im[pos_idxs[is_in_gts == 0]] = -1
else:
raise NotImplementedError
reg_targets_per_im = self.box_coder.encode(matched_gts, anchors_per_im.bbox)
cls_labels.append(cls_labels_per_im)
reg_targets.append(reg_targets_per_im)
return cls_labels, reg_targets
