def select_bg(self, Phi_labels, boxes, labels, bbox_pred, keeps_Y, good_gt_overlap, M, im_shape_w, im_shape_h):
"""
Find B in p(B|Xb)
"""
selected_item = range(M)
prob_dpp = np.ones((M,))
ignores = []
dict_keeps_Y = {}
for i, j in keeps_Y.iteritems():
if j not in dict_keeps_Y:
dict_keeps_Y[j] = []
dict_keeps_Y[j].append(i)
for k in range(M):
if (k in keeps_Y and keeps_Y[k] == Phi_labels[k]) \
or (k in good_gt_overlap and Phi_labels[k] == labels[k] and labels[k] > 0):
ignores.append(k)
else:
label_k = labels[k]
if label_k in dict_keeps_Y:
loc_lbl = bbox_pred[[k], 4 * label_k:4 * (label_k + 1)]
pbox = bbox_transform_inv(boxes[[k], :], loc_lbl)
pbox = clip_boxes(pbox, (im_shape_w, im_shape_h))
pbox = np.reshape(np.tile(pbox, len(dict_keeps_Y[label_k])), (len(dict_keeps_Y[label_k]), 4))
Y_selected_ll = bbox_pred[dict_keeps_Y[label_k], 4 * label_k:4 * (label_k + 1)]
Y_selected_pbox = bbox_transform_inv(boxes[dict_keeps_Y[label_k], :], Y_selected_ll)
Y_selected_pbox = clip_boxes(Y_selected_pbox, (im_shape_w, im_shape_h))
if np.max(IoU_target(pbox, Y_selected_pbox)) > cfg.TRAIN.IGNORANCE:
ignores.append(k)
selected_item = np.array([x for ii, x in enumerate(selected_item) if ii not in ignores])
prob_dpp = [x for ii, x in enumerate(prob_dpp) if ii not in ignores]
return selected_item, prob_dpp
def update_rl(rl_in, h_start, w_start, h_end, w_end, t, rois_seq,
cls_probs_seq, bbox_preds_seq, cls_probs_uptonow,
pred_bboxes_uptonow, keeps, im_shape, bin_ctrs, height,
width, rl_in_upsamp_height, rl_in_upsamp_width, thresh=0.0):
if t > 1:
cls_probs_uptonow = cls_probs_uptonow[keeps[0], :]
pred_bboxes_uptonow = pred_bboxes_uptonow[keeps[0], :]
keeps[0] = []
# Potentially perform per-time-step NMS
if rois_seq is not None:
# Current preds at this fix merged with survivors from previous steps
cls_probs_uptonow = np.vstack([cls_probs_uptonow, cls_probs_seq])
pred_bboxes = bbox_transform_inv(rois_seq, bbox_preds_seq)
pred_bboxes = clip_boxes(pred_bboxes, im_shape)
pred_bboxes_uptonow = np.vstack([pred_bboxes_uptonow, pred_bboxes])
# Perform on-the-fly NMS (used when performing class-specific history updates)
keeps = _get_nms_keep(keeps, cls_probs_uptonow, pred_bboxes_uptonow, thresh)
# Update non-history part of RL state
rl_in[:, h_start:h_end, w_start:w_end, :cfg.DIMS_NONHIST] = -1
if rois_seq is not None:
rl_in, _ = do_hist_update(rl_in, cls_probs_uptonow, pred_bboxes_uptonow,
keeps, bin_ctrs, height, width,
rl_in_upsamp_height, rl_in_upsamp_width)
return rl_in, keeps, cls_probs_uptonow, pred_bboxes_uptonow
def proposal_top_layer(rpn_cls_prob, rpn_bbox_pred, im_info, _feat_stride, anchors, num_anchors, rpn_reject_inds):
"""A layer that just selects the top region proposals
without using non-maximal suppression,
For details please see the technical report
"""
rpn_top_n = cfg.TEST.RPN_TOP_N
im_info = im_info[0]
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
length = scores.shape[0]
######################REJECT VIA RPN################
###------------------------reject process---------------------------###
if rpn_reject_inds.size != 0:
reject_inds = np.unique(rpn_reject_inds)
scores[reject_inds] = -2
passinds = np.where(scores != -2)[0]
#reject via frcn and rpn
anchors = anchors[passinds]
scores = scores[passinds]
rpn_bbox_pred = rpn_bbox_pred[passinds]
###-------------------------reject done-----------------------------###
#####################################################
length = scores.shape[0]
if length < rpn_top_n:
# Random selection, maybe unnecessary and loses good proposals
# But such case rarely happens
top_inds = npr.choice(length, size=rpn_top_n, replace=True)
else:
top_inds = scores.argsort(0)[::-1]
top_inds = top_inds[:rpn_top_n]
top_inds = top_inds.reshape(rpn_top_n, )
# Do the selection here
anchors = anchors[top_inds, :]
rpn_bbox_pred = rpn_bbox_pred[top_inds, :]
scores = scores[top_inds]
# Convert anchors into proposals via bbox transformations
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
# Clip predicted boxes to image
proposals = clip_boxes(proposals, im_info[:2])
# Output rois blob
# Our RPN implementation only supports a single input image, so all
# batch inds are 0
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def pad_rois(rois, im_info, is_training):
"""Pad rois to utilize contextual information and to alleviate truncation
"""
nroi = rois.shape[0]
proposals = np.zeros((nroi, 4), dtype=np.float32)
w = rois[:, 3] - rois[:, 1]
h = rois[:, 4] - rois[:, 2]
dw = cfg.POOL_PAD_RATIO * w
dh = cfg.POOL_PAD_RATIO * h
nroi = rois.shape[0]
if is_training:
nw = npr.rand(nroi)
nh = npr.rand(nroi)
else:
nw = np.ones(nroi) * 0.5
nh = np.ones(nroi) * 0.5
proposals[:, 0] = rois[:, 1] - (dw - nw *
(1 + 2 * cfg.POOL_PAD_RATIO) / 15 * w)
proposals[:, 1] = rois[:, 2] - (dh - nh *
(1 + 2 * cfg.POOL_PAD_RATIO) / 15 * h)
proposals[:, 2] = rois[:, 3] + (dw - (1 - nw) *
(1 + 2 * cfg.POOL_PAD_RATIO) / 15 * w)
proposals[:, 3] = rois[:, 4] + (dh - (1 - nh) *
(1 + 2 * cfg.POOL_PAD_RATIO) / 15 * h)
proposals = clip_boxes(proposals, im_info[:2])
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride, anchors, num_anchors):
'''
A simplified version compared to fast/er RCNN
For details please see the technical report
:param rpn_cls_prob:
:param rpn_bbox_pred:
:param im_info: [M,N,scale_factor]保存了将任意图像缩放到M×N的所有信息
:param cfg_key:
:param _feat_stride:feat_stride=16用于计算anchor的偏移量
:param anchors:
:param num_anchors:
:return:
'''
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
# 计算得到bbox四个顶点坐标
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
'''
按照输入的foreground softmax降序排列，提取前pre_nms_topN（6000）的结果
提取修正后的foreground anchor
'''
order = scores.ravel().argsort()[::-1]
# ravel数组扁平化，降序排列
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
proposals = proposals[order, :]
# anchor坐标
scores = scores[order]
# anchor分数
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores)), nms_thresh)
# Pick th top region proposals after NMS
'''
再次按照nms后的foreground softmax由大到小排列，提取前post_nms_topN（300）结果作为proposals的输出
'''
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep]
# Only support single image as input
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors, reject_inds_1, reject_inds_2):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
im_info = im_info[0]
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
######################REJECT VIA RPN################
###------------------------reject process---------------------------###
if reject_inds_1.size != 0:
reject_inds_1 = np.unique(reject_inds_1)
scores[reject_inds_1] = -2
if reject_inds_2.size != 0:
reject_inds_2 = np.unique(reject_inds_2)
scores[reject_inds_2] = -2
passinds = np.where(scores != -2)[0]
#reject via frcn and rpn
proposals = proposals[passinds]
scores = scores[passinds]
###-------------------------reject done-----------------------------###
#####################################################
# Pick the top region proposals
order = scores.ravel().argsort()[::-1]
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
proposals = proposals[order, :]
scores = scores[order]
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores)), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep]
# Only support single image as input
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
# cfg_key代表TRAIN还是TEST
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
'''
pre_nms_topN: 在NMS处理之前，分数在前面的rois
post_nms_topN: 在NMS处理之后，分数在前面的rois
nms_thresh: NMS的阈值
'''
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH # __C.TRAIN.RPN_NMS_THRESH = 0.7
# Get the scores and bounding boxes
# 其中第四维度前9位是背景的分数，后9位是前景的分数
# 假设rpn_cls_prob = (1,38,50,18)
scores = rpn_cls_prob[:, :, :, num_anchors:] # scores = (1,38,50,9)
rpn_bbox_pred = rpn_bbox_pred.reshape(
(-1, 4)) # rpn_bbox_pred = （1,38,50,36）->(17100,4)
scores = scores.reshape((-1, 1)) # scores = (17100,1)
# bbox_transform_inv 根据anchor和偏移量计算proposals
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
# clip_boxes作用：调整boxes的坐标，使其全部在图像的范围内
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
# 首先变成一维，然后argsort返回数组值从小到大的索引值,然后加上[::-1]，翻转序列
# order保存数组值从大到小的索引值
order = scores.ravel().argsort()[::-1]
if pre_nms_topN > 0: # 只取前pre_nms_topN
order = order[:pre_nms_topN]
# order对应的是下标，然后把得分最高的前pre_nms_topN的区域保存
proposals = proposals[order, :]
# 只保存前pre_nms_topN个得分
scores = scores[order]
# Non-maximal suppression
# 非极大值抑制 np.hstack把他们拼接成(区域 分数)的形式
keep = nms(np.hstack((proposals, scores)), nms_thresh)
# Pick th top region proposals after NMS
# 在nms之后，选择前post_nms_topN个
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep]
# Only support single image as input
# 这点多出来一个batch_inds，拼接之后blob的第一列全是0，不知道后面是不是有什么操作。。。
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors, gt_boxes, gt_texts, gt_pair):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.view((-1, 4))
scores = scores.contiguous().view(-1, 1)
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
scores, order = scores.view(-1).sort(descending=True)
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
scores = scores[:pre_nms_topN].view(-1, 1)
proposals = proposals[order.data, :]
# Till now, proposals and scores consists RPN_PRE_NMS_TOP_N (12000/6000) top anchor_regions
# only by the rpn_cls_prob
'''
if cfg.mode=='TRAIN':
overlaps = bbox_overlaps(
proposals.data,
gt_boxes[:, :4].data)
max_overlaps, gt_assignment = overlaps.max(1)
labels = gt_boxes[gt_assignment, [4]]
texts = [gt_texts[i] for i in gt_assignment]
pair = torch.LongTensor([int(gt_pair[i]) for i in gt_assignment]).cuda()
print("labels")
print(labels)
print("pair")
print(pair)
'''
# Non-maximal suppression
keep = nms(torch.cat((proposals, scores), 1).data, nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep, ]
# Only support single image as input
batch_inds = Variable(proposals.data.new(proposals.size(0), 1).zero_())
blob = torch.cat((batch_inds, proposals), 1)
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
blob = []
scores = []
# Get the scores and bounding boxes
for im_i in np.arange(im_info.shape[0]):
scores_im_i = rpn_cls_prob[im_i, :, :, num_anchors:].copy().reshape(
(-1, 1))
rpn_bbox_pred_im_i = rpn_bbox_pred[im_i].copy().reshape((-1, 4))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred_im_i)
proposals = clip_boxes(proposals, im_info[im_i, :2])
# Pick the top region proposals
order = scores_im_i.ravel().argsort()[::-1]
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
proposals = proposals[order, :]
scores_im_i = scores_im_i[order]
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores_im_i)), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
# not filter out the remaining, but re-ranking and then clip
scores_im_i[keep] += 2
order = scores_im_i.ravel().argsort()[::-1]
keep = order[:post_nms_topN]
proposals = proposals[keep, :]
scores_im_i = scores_im_i[keep]
scores.append(scores_im_i)
# multi image as input
batch_inds = im_i * np.ones((proposals.shape[0], 1), dtype=np.float32)
blob_im_i = np.hstack(
(batch_inds, proposals.astype(np.float32, copy=False)))
blob.append(blob_im_i)
scores = np.concatenate(scores)
blob = np.concatenate(blob)
return blob, scores
def unnormalize_box(Phi_labels, bbox_targets, boxes, stds, means,M, im_shape_w, im_shape_h):
"""
un-normalize boxes by using stds and means
"""
Phi_argmax = 4 * Phi_labels
bbox_target = bbox_targets[np.tile(range(M),4),np.hstack((4*Phi_labels,4*Phi_labels+1,4*Phi_labels+2, 4*Phi_labels+3))]
bbox_target = np.reshape(bbox_target,(M,4),order='F')
bbox_target = bbox_target * stds[Phi_argmax/4,:] + means[Phi_argmax/4,:]
unnormalized_bbox_targets = bbox_transform_inv(boxes, bbox_target)
unnormalized_bbox_targets = clip_boxes(unnormalized_bbox_targets, (im_shape_w,im_shape_h))
return unnormalized_bbox_targets
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
Parameters:
- rpn_cls_prob : (1, H, W, 2A) float Variable
- rpn_bbox_pred: (1, H, W, 4A)
- im_info : [im_height, im_width, scale], ndarray (3, )
- cfg_key : train or test
- _feat_stride : 16
- anchors : (HWA, 4) float Variable
- num_anchors : A = 9
Returns:
- blob : Variable (N_nms, 5) [0; x1y1x2h2]
- scores : Variable (N_nms, )
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N # train 12000; test 6000
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N # train 2000 ; test 300
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
im_info = im_info[0]
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:] # (1, H, W, A) pos score only
rpn_bbox_pred = rpn_bbox_pred.view((-1, 4)) # (HWA, 4)
scores = scores.contiguous().view(-1, 1) # (HWA, 1)
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2]) # (HWA, 4)
# Pick the top region proposals
scores, order = scores.view(-1).sort(descending=True)
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
scores = scores[:pre_nms_topN].view(-1, 1)
proposals = proposals[order.data, :]
# Non-maximal suppression
# changes by me
keep = nms(proposals, scores[:, 0], nms_thresh)
# keep = nms(torch.cat((proposals, scores), 1).data, nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep, ]
# Only support single image as input
batch_inds = Variable(proposals.data.new(proposals.size(0), 1).zero_())
blob = torch.cat((batch_inds, proposals), 1)
return blob, scores
def forward(self, out_cls, out_ellipse):
"""
out_cls: (feat_height, feat_width, anchors, 2) FloatVariable
out_ellipse: (feat_height, feat_width, anchors, 5) FloatVariable
"""
scores = nn.functional.softmax(out_cls,
dim=3)[...,
1].contiguous().data.view(-1, 1)
ellipse_deltas = out_ellipse.data.view(-1, 5)
# 1. Generate proposals from ellipse deltas and shifted anchors
# Convert anchors into proposals via ellipse transformations
# Convert ellipse into bbox proposals
ellipses = ellipse_transform_inv(self._anchors, ellipse_deltas)
boxes = ellipse2box(ellipses, self._cfg['ELLIPSE_PAD'])
# 2. clip predicted boxes to image
boxes = clip_boxes(boxes, self._im_info[:2])
# 3. remove predicted boxes with either height or width < threshold
# (NOTICE: convert min_size to input image scale stored in im_info[2])
keep = _filter_boxes(boxes, self._cfg['TEST.RPN_MIN_SIZE'])
boxes = boxes[keep, :]
ellipses = ellipses[keep, :]
scores = scores[keep]
# 4. sort all (proposal, score) pairs by score from highest to lowest
# 5. take top pre_nms_topN (e.g. 6000)
_, order = torch.sort(scores.view(-1), dim=0, descending=True)
if self._cfg['TEST.RPN_PRE_NMS_TOP_N'] > 0:
order = order[:self._cfg['TEST.RPN_PRE_NMS_TOP_N']]
boxes = boxes[order, :]
ellipses = ellipses[order, :]
scores = scores[order]
# 6. apply nms (e.g. threshold = 0.7)
# 7. take after_nms_topN (e.g. 300)
# 8. return the top proposals (-> RoIs top)
if self._cfg['USE_GPU_NMS']:
nms = gpu_nms
else:
nms = cpu_nms
dets = np.hstack((boxes.cpu().numpy(), scores.cpu().numpy()))
keep = nms(dets, self._cfg['TEST.RPN_NMS_THRESH'])
keep = torch.from_numpy(np.array(keep)).type_as(scores).long()
if self._cfg['TEST.RPN_POST_NMS_TOP_N'] > 0:
keep = keep[:self._cfg['TEST.RPN_POST_NMS_TOP_N']]
boxes = boxes[keep, :]
ellipses = ellipses[keep, :]
scores = scores[keep].view(-1)
return (boxes, ellipses, scores)
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, anchors,
num_anchors):
"""
A simplified version compared to fast/er RCNN
For details please see the technical report
:param
rpn_cls_prob: (1, H, W, Ax2) softmax result of rpn scores
rpn_bbox_pred: (1, H, W, Ax4) 1x1 conv result for rpn bbox
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
# Get the scores and bounding boxes for foreground (text)
# The order in last dim is related to network.py:
# self._reshape_layer(rpn_cls_prob_reshape, self._num_anchors * 2, "rpn_cls_prob")
# scores = rpn_cls_prob[:, :, :, num_anchors:] # old
height, width = rpn_cls_prob.shape[1:3] # feature-map的高宽
scores = np.reshape(
np.reshape(rpn_cls_prob,
[1, height, width, num_anchors, 2])[:, :, :, :, 1],
[1, height, width, num_anchors])
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
order = scores.ravel().argsort()[::-1]
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
proposals = proposals[order, :]
scores = scores[order]
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores)), nms_thresh, not cfg.USE_GPU_NMS)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep]
# Only support single image as input
blob = np.hstack(
(scores.astype(np.float32,
copy=False), proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_top_layer(rpn_cls_prob, rpn_bbox_pred, im_info, _feat_stride,
anchors, num_anchors):
"""A layer that just selects the top region proposals
without using non-maximal suppression,
For details please see the technical report
Parameters:
- rpn_cls_prob : (1, H, W, 2A) float Variable
- rpn_bbox_pred: (1, H, W, 4A)
- im_info : [im_height, im_width, scale], ndarray (3, )
- _feat_stride : 16
- anchors : (HWA, 4) float Variable
- num_anchors : A = 9
Returns:
- blob : (N_nms, 5) float Variable [0; x1y1x2h2]
- scores : (N_nms, ) float Variable
"""
rpn_top_n = cfg.TEST.RPN_TOP_N
im_info = im_info[0]
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.view(-1, 4)
scores = scores.contiguous().view(-1, 1)
length = scores.size(0)
if length < rpn_top_n:
# Random selection, maybe unnecessary and loses good proposals
# But such case rarely happens
top_inds = torch.from_numpy(
npr.choice(length, size=rpn_top_n, replace=True)).long().cuda()
else:
top_inds = scores.sort(0, descending=True)[1]
top_inds = top_inds[:rpn_top_n]
top_inds = top_inds.view(rpn_top_n)
# Do the selection here
anchors = anchors[top_inds].contiguous()
rpn_bbox_pred = rpn_bbox_pred[top_inds].contiguous()
scores = scores[top_inds].contiguous()
# Convert anchors into proposals via bbox transformations
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
# Clip predicted boxes to image
proposals = clip_boxes(proposals, im_info[:2])
# Output rois blob
# Our RPN implementation only supports a single input image, so all
# batch inds are 0
batch_inds = Variable(proposals.data.new(proposals.size(0), 1).zero_())
blob = torch.cat([batch_inds, proposals], 1)
return blob, scores
def proposal_top_layer(rpn_cls_prob, rpn_bbox_pred, im_info, _feat_stride, anchors, num_anchors):
"""A layer that just selects the top region proposals
without using non-maximal suppression,
For details please see the technical report
"""
# __C.TEST.RPN_TOP_N = 5000 仅TEST.MODE = 'top' 的时候使用
# __C.TEST.MODE = 'nms'
rpn_top_n = cfg.TEST.RPN_TOP_N
# 提取概率分数
scores = rpn_cls_prob[:, :, :, num_anchors:]
# 对提取的预测狂reshape # rpn_bbox_pred：RPN层输出的box的取值，即：tx,ty,tw,th
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
# 统计有多少个框
length = scores.shape[0]
if length < rpn_top_n: # 如果框小于5000个，需要随即重复采样，让他变成5000个
# Random selection, maybe unnecessary and loses good proposals
# But such case rarely happens
top_inds = npr.choice(length, size=rpn_top_n, replace=True)
else:
# 从大到小排序，取列索引
top_inds = scores.argsort(0)[::-1]
# 取前大的5000个
top_inds = top_inds[:rpn_top_n]
top_inds = top_inds.reshape(rpn_top_n, )
# Do the selection here
# 选择/重排
# 按照索引提取anchor数据
anchors = anchors[top_inds, :]
rpn_bbox_pred = rpn_bbox_pred[top_inds, :]
scores = scores[top_inds]
# Convert anchors into proposals via bbox transformations
# bbox_transform_inv : 根据anchor和偏移量计算proposals
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
# Clip predicted boxes to image
# clip_boxes : proposals的边界限制在图片内
proposals = clip_boxes(proposals, im_info[:2])
# Output rois blob
# Our RPN implementation only supports a single input image, so all
# batch inds are 0
# 和 proposal_layer 一样，多出来一列0，然后拼接
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
#按照通道C取出RPN预测的框属于前景的分数，请注意，这18个channel中，前9个是背景的概率，后九个才是前景的概率
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
proposals = bbox_transform_inv(
anchors, rpn_bbox_pred) #在这里结合RPN的输出变换初始框的坐标，得到第一次变换坐标后的proposals
proposals = clip_boxes(proposals,
im_info[:2]) #在这里讲超出图像边界的proposal进行边界裁剪，使之在图像边界之内
# 按照前景概率进行排序，取前top个，
#对框按照前景分数进行排序，order中指示了框的下标
# Pick the top region proposals
order = scores.ravel().argsort()[::-1]
if pre_nms_topN > 0:
order = order[:
pre_nms_topN] #选择前景分数排名在前pre_nms_topN(训练时为12000，测试时为6000)的框
proposals = proposals[order, :] #保留了前pre_nms_topN个框的坐标信息
scores = scores[order] #保留了前pre_nms_topN个框的分数信息
# 对剩下的proposal进行NMS操作，阈值是0.7进行nms操作，再取前n个
# Non-maximal suppression
#使用nms算法排除重复的框
keep = nms(np.hstack((proposals, scores)), nms_thresh)
# 对剩下的proposal，保留RPN_POST_NMS_TOP_N个， 得到最终的rois和相应的rpn_socre
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:
post_nms_topN] #选择前景分数排名在前post_nms_topN(训练时为2000，测试时为300)的框
proposals = proposals[keep, :] #保留了前post_nms_topN个框的坐标信息
scores = scores[keep] #保留了前post_nms_topN个框的分数信息
# Only support single image as input
# 因为要进行roi_pooling，所以在保留框内的坐标信息前面插入batch中图片的编号信息，此时，batchsize为1，所以都插入为0
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_top_layer(rpn_cls_prob, rpn_bbox_pred, im_info, _feat_stride,
anchors, num_anchors):
"""A layer that just selects the top region proposals
without using non-maximal suppression,
For details please see the technical report
"""
rpn_top_n = cfg.TEST.RPN_TOP_N
blob = []
scores = []
# Get the scores and bounding boxes
for im_i in np.arange(im_info.shape[0]):
scores_ind_start = num_anchors if cfg.TRAIN.RPN_FL_SOFTMAX else 0
scores_im_i = rpn_cls_prob[im_i, :, :,
scores_ind_start:].copy().reshape((-1, 1))
rpn_bbox_pred_im_i = rpn_bbox_pred[im_i].copy().reshape((-1, 4))
length = scores_im_i.shape[0]
if length < rpn_top_n:
# Random selection, maybe unnecessary and loses good proposals
# But such case rarely happens
top_inds = npr.choice(length, size=rpn_top_n, replace=True)
else:
top_inds = scores_im_i.argsort(0)[::-1]
top_inds = top_inds[:rpn_top_n]
top_inds = top_inds.reshape(rpn_top_n, )
# Do the selection here
anchors = anchors[top_inds, :]
rpn_bbox_pred_im_i = rpn_bbox_pred_im_i[top_inds, :]
scores_im_i = scores_im_i[top_inds]
scores.append(scores_im_i)
# Convert anchors into proposals via bbox transformations
proposals = bbox_transform_inv(anchors, rpn_bbox_pred_im_i)
# Clip predicted boxes to image
proposals = clip_boxes(proposals, im_info[im_i, :2])
# multi image as input
batch_inds = im_i * np.ones((proposals.shape[0], 1), dtype=np.float32)
blob_im_i = np.hstack(
(batch_inds, proposals.astype(np.float32, copy=False)))
blob.append(blob_im_i)
scores = np.concatenate(scores)
blob = np.concatenate(blob)
return blob, scores
def proposal_layer_test_caption_compact(rpn_cls_prob, rpn_bbox_pred, im_info,
cfg_key, _feat_stride, anchors,
num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
order = scores.ravel().argsort()[::-1]
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
proposals = proposals[order, :]
scores = scores[order]
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores)), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep]
# Only support single image as input
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
# if cfg.DEBUG_VERBOSE == 1:
# print('_region_proposal.py ~~~~~~~~~~~~~ TESTING ... blob shape {:s}'.format(blob.shape)) # e.g. (210, 5)
# print('_region_proposal.py ~~~~~~~~~~~~~ TESTING ... sentences shape {:s}'.format(sentences.shape)) # e.g. (210, 10)
# print('_region_proposal.py ~~~~~~~~~~~~~ TESTING ... sentences {:s}'.format(sentences))
# print('_region_proposal.py ~~~~~~~~~~~~~ TESTING ... score shape {:s}'.format(scores.shape))
# print('_region_proposal.py ~~~~~~~~~~~~~ TESTING ... blob (rois) {:s}'.format(blob))
# print('_region_proposal.py ~~~~~~~~~~~~~ TESTING ... score {:s}'.format(scores))
return blob, scores
def proposal_layer(rpn_cls_prob,
rpn_bbox_pred,
im_info,
cfg_key,
_feat_stride,
anchors,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
scales = np.array(anchor_scales)
ratios = np.array(anchor_ratios)
num_anchors = scales.shape[0] * ratios.shape[0]
im_info = im_info[0]
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
order = scores.ravel().argsort()[::-1]
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
proposals = proposals[order, :]
scores = scores[order]
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores)), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep]
# Only support single image as input
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_layer_fpn(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key,
_feat_stride, anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
proposals_total = []
scores_total = []
for idx in range(len(rpn_cls_prob)):
# Get the scores and bounding boxes
scores = rpn_cls_prob[idx][:, :, :, num_anchors:]
rpn_bbox_pred[idx] = rpn_bbox_pred[idx].view((-1, 4))
scores = scores.contiguous().view(-1, 1)
proposals = bbox_transform_inv(anchors[idx], rpn_bbox_pred[idx])
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
scores, order = scores.view(-1).sort(descending=True)
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
scores = scores[:pre_nms_topN].view(-1, 1)
proposals = proposals[order.data, :]
proposals_total.append(proposals)
scores_total.append(scores)
proposals = torch.cat(proposals_total)
scores = torch.cat(scores_total)
# Non-maximal suppression
keep = nms(torch.cat((proposals, scores), 1).data, nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep, ]
# Only support single image as input
batch_inds = Variable(proposals.data.new(proposals.size(0), 1).zero_())
blob = torch.cat((batch_inds, proposals), 1)
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""
A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N #12000
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N #2000
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH #0.7
# Get the scores and bounding boxes
# [1,57,38,18] [1,57,38,36]
scores = rpn_cls_prob[:, :, :, num_anchors:] #[1,57,38,9]
rpn_bbox_pred = rpn_bbox_pred.view((-1, 4)) #[19494,4]
scores = scores.contiguous().view(-1, 1) # [19494,1]
#9个anchor，[19494,4]，做边框平移和缩放，得到和ground truth相近的结果
proposals = bbox_transform_inv(anchors, rpn_bbox_pred) #[x1,y1,x2,y2]
proposals = clip_boxes(proposals, im_info[:2]) #根据宽高缩放比，w，h，scale
# Pick the top region proposals
# scores是值，order是序列
scores, order = scores.view(-1).sort(descending=True)
#保留2000个最高的
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
scores = scores[:pre_nms_topN].view(-1, 1)
proposals = proposals[order.data, :]
# Non-maximal suppression
keep = nms(torch.cat((proposals, scores), 1).data, nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep, ]
# Only support single image as input #保留2000个最高的
batch_inds = proposals.new_zeros(proposals.size(0), 1) #[2000,1]，全0
blob = torch.cat((batch_inds, proposals), 1) #[2000,5] 多一维batch_inds
return blob, scores #[2000,5] [x1,y1,x2,y2] [2000,1]
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
min_size = cfg[cfg_key].ANCHOR_MIN_SIZE
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
# removed predicted boxes with either height or width < threshold
# (NOTE: convert min_size to input image scale stored in im_info[2])
# keep = _filter_boxes(proposals, min_size * im_info[2])
# proposals = proposals[keep, :]
# scores = scores[keep]
# Pick the top region proposals
order = scores.ravel().argsort()[::-1]
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
proposals = proposals[order, :]
scores = scores[order]
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores)), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep]
# Only support single image as input
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.view((-1, 4))
scores = scores.contiguous().view(-1, 1)
# print('anchors: ', anchors)
# print('rpn_bbox_pred: ', rpn_bbox_pred)
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
# print('PROPOSALS0: ', proposals)
proposals = clip_boxes(proposals, im_info[:2])
# print('PROPOSALS1: ', proposals)
# Pick the top region proposals
scores, order = scores.view(-1).sort(descending=True)
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
scores = scores[:pre_nms_topN].view(-1, 1)
proposals = proposals[order.data, :]
# Non-maximal suppression
keep = nms(proposals, scores.squeeze(1), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep, ]
# Only support single image as input
batch_inds = proposals.new_zeros(proposals.size(0), 1)
blob = torch.cat((batch_inds, proposals), 1)
# print('PROPOSALS: ', proposals)
# print("BLOOOOOB: ", blob)
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
#12000 anchors pre nms algorithm are applied!
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
#2000 anchors are generated from one pic which are generated after nms
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
#the threshold is set to 0.7 to delete abundant anchors!
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
#
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
order = scores.ravel().argsort()[::-1]
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
proposals = proposals[order, :]
scores = scores[order]
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores)), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep]
# Only support single image as input
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_top_layer(rpn_cls_prob, rpn_bbox_pred, im_info, _feat_stride,
anchors, num_anchors):
"""A layer that just selects the top region proposals
without using non-maximal suppression,
For details please see the technical report
"""
rpn_top_n = cfg.TEST.RPN_TOP_N
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.view(-1, 4)
scores = scores.contiguous().view(-1, 1)
length = scores.size(0)
if length < rpn_top_n:
# Random selection, maybe unnecessary and loses good proposals
# But such case rarely happens
top_inds = torch.from_numpy(
npr.choice(length, size=rpn_top_n, replace=True)).long().cuda()
else:
top_inds = scores.sort(0, descending=True)[1]
top_inds = top_inds[:rpn_top_n]
top_inds = top_inds.view(rpn_top_n)
# Do the selection here
anchors = anchors[top_inds, :].contiguous()
rpn_bbox_pred = rpn_bbox_pred[top_inds, :].contiguous()
scores = scores[top_inds].contiguous()
# Convert anchors into proposals via bbox transformations
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
# Clip predicted boxes to image
proposals = clip_boxes(proposals, im_info[:2])
# Output rois blob
# Our RPN implementation only supports a single input image, so all
# batch inds are 0
batch_inds = proposals.data.new(proposals.size(0), 1).zero_()
# blob = torch.cat([batch_inds, proposals], 1)
## Changed to make __C.TEST.MODE = 'top' work --> blob = torch.cat((Variable(batch_inds), proposals), 1)
blob = torch.cat((Variable(batch_inds), proposals), 1)
return blob, scores
def proposal_layer_all(rpn_bbox_pred, im_info, anchors, rpn_cls_prob=None):
"""
Simply returns every single RoI; drl-RPN later decides
which are forwarded to the class-specific module.
"""
# Get the bounding boxes
batch_sz, height, width = rpn_bbox_pred.shape[0:3]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
# Create initial (all-zeros) observation RoI volume
roi_obs_vol = np.zeros((batch_sz, height, width, cfg.NBR_ANCHORS),
dtype=np.int32)
if cfg.DRL_RPN.USE_AGNO:
# If this branch is used, we only consider RoIs among survivors from
# class-agnositc NMS when choosing RoIs with drl-RPN
pre_nms_topN = cfg.TEST.RPN_PRE_NMS_TOP_N
post_nms_topN = cfg.TEST.RPN_POST_NMS_TOP_N
nms_thresh = cfg.TEST.RPN_NMS_THRESH
scores = rpn_cls_prob[:, :, :, cfg.NBR_ANCHORS:]
scores = scores.reshape((-1, 1))
keep_ids_all = np.arange(scores.shape[0], dtype=np.int32)
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores)), nms_thresh)
keep_ids = keep_ids_all[keep]
not_keep_ids = np.setdiff1d(keep_ids_all, keep_ids)
else:
not_keep_ids = np.zeros((1, 1), dtype=np.int32)
# Only support single image as input
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
rois_all = np.hstack((batch_inds, proposals.astype(np.float32,
copy=False)))
return rois_all, roi_obs_vol, not_keep_ids
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, info, cfg_key, anchors,
anchors_3d, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
#print('rpn_bbox_pred')
#print(rpn_bbox_pred)
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :,
num_anchors:] #Start at back half of anchor/score list
rpn_bbox_pred = rpn_bbox_pred.view(
(-1, 4)) #rpn_bbox_pred are adjustment factors to existing anchors
scores = scores.contiguous().view(-1, 1) #Collapse into a single vector
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, info) #Make sure they are within bounds
# Pick the top 'pre_nms_topN' # of region proposals
scores, order = scores.view(-1).sort(descending=True)
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
scores = scores[:pre_nms_topN].view(-1, 1)
proposals = proposals[order.data, :]
anchors_3d = anchors_3d[order.data, :]
# Non-maximal suppression
keep = nms(proposals, scores.squeeze(1), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep, ]
anchors_3d = anchors_3d[keep, :]
# Only support single image as input
batch_inds = proposals.new_zeros(proposals.size(0), 1)
blob = torch.cat((batch_inds, proposals), 1)
return blob, scores, anchors_3d
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""
1,根据rpn_bbox_pred对anchor进行微调
2,根据预测分数选前pre_nms_topN个anchor
3,进行nms非极大值抑制处理
4,对非极大值抑制处理后的数据,取前post_nms_topN个推荐框
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
order = scores.ravel().argsort()[::-1]
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
proposals = proposals[order, :]
scores = scores[order]
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores)), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep]
# Only support single image as input
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_top_layer(rpn_cls_prob, rpn_bbox_pred, im_info, _feat_stride,
anchors, num_anchors):
"""A layer that just selects the top region proposals
without using non-maximal suppression,
For details please see the technical report
"""
rpn_top_n = cfg.TEST.RPN_TOP_N
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.view(-1, 4)
scores = scores.contiguous().view(-1, 1)
length = scores.size(0)
if length < rpn_top_n:
# Random selection, maybe unnecessary and loses good proposals
# But such case rarely happens
top_inds = torch.from_numpy(
npr.choice(length, size=rpn_top_n,
replace=True)).long().to(anchors.device)
else:
top_inds = scores.sort(0, descending=True)[1]
top_inds = top_inds[:rpn_top_n]
top_inds = top_inds.view(rpn_top_n)
# Do the selection here
anchors = anchors[top_inds, :].contiguous()
rpn_bbox_pred = rpn_bbox_pred[top_inds, :].contiguous()
scores = scores[top_inds].contiguous()
# Convert anchors into proposals via bbox transformations
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
# Clip predicted boxes to image
proposals = clip_boxes(proposals, im_info[:2])
# Output rois blob
# Our RPN implementation only supports a single input image, so all
# batch inds are 0
batch_inds = proposals.new_zeros(proposals.size(0), 1)
blob = torch.cat([batch_inds, proposals], 1)
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.view((-1, 4))
scores = scores.contiguous().view(-1, 1)
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
scores, order = scores.view(-1).sort(descending=True)
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
scores = scores[:pre_nms_topN].view(-1, 1)
proposals = proposals[order.data, :]
# Non-maximal suppression
keep = nms(torch.cat((proposals, scores), 1).data, nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :] # test(300,4)
scores = scores[keep, ]
# Our RPN implementation only supports a single input image,
# so all batch inds are 0
# 即这些roi都属于一个图片，如果后续实现了多个输入图片，这个roi要区分它属于哪一个图片(即哪一个batch)
batch_inds = Variable(proposals.data.new(proposals.size(0), 1).zero_())
blob = torch.cat((batch_inds, proposals), 1)
return blob, scores
def predict(self, blobs):
"""
Test image
:param im: bgr
:return:
"""
# Test
img_h, img_w, im_scale = blobs['im_info']
_, scores, bbox_pred, rois = self.net.test_image(
blobs['data'], blobs['im_info'])
boxes = rois[:, 1:5] / im_scale
scores = np.reshape(scores, [scores.shape[0], -1])
bbox_pred = np.reshape(bbox_pred, [bbox_pred.shape[0], -1])
# Apply bounding-box regression deltas
box_deltas = bbox_pred
pred_boxes = bbox_transform_inv(t.from_numpy(boxes),
t.from_numpy(box_deltas))
pred_boxes = clip_boxes(pred_boxes, [img_h, img_w]).numpy()
return scores, pred_boxes
def proposal_top_layer(rpn_cls_prob, rpn_bbox_pred, im_info, _feat_stride, anchors, num_anchors):
"""A layer that just selects the top region proposals
without using non-maximal suppression,
For details please see the technical report
"""
rpn_top_n = cfg.TEST.RPN_TOP_N
im_info = im_info[0]
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
length = scores.shape[0]
if length < rpn_top_n:
# Random selection, maybe unnecessary and loses good proposals
# But such case rarely happens
top_inds = npr.choice(length, size=rpn_top_n, replace=True)
else:
top_inds = scores.argsort(0)[::-1]
top_inds = top_inds[:rpn_top_n]
top_inds = top_inds.reshape(rpn_top_n, )
# Do the selection here
anchors = anchors[top_inds, :]
rpn_bbox_pred = rpn_bbox_pred[top_inds, :]
scores = scores[top_inds]
# Convert anchors into proposals via bbox transformations
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
# Clip predicted boxes to image
proposals = clip_boxes(proposals, im_info[:2])
# Output rois blob
# Our RPN implementation only supports a single input image, so all
# batch inds are 0
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride, anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
im_info = im_info[0]
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.reshape((-1, 4))
scores = scores.reshape((-1, 1))
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
order = scores.ravel().argsort()[::-1]
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
proposals = proposals[order, :]
scores = scores[order]
# Non-maximal suppression
keep = nms(np.hstack((proposals, scores)), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep]
# Only support single image as input
batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)
blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))
return blob, scores
def proposal_layer(rpn_cls_prob, rpn_bbox_pred, im_info, cfg_key, _feat_stride,
anchors, num_anchors):
"""A simplified version compared to fast/er RCNN
For details please see the technical report
"""
if type(cfg_key) == bytes:
cfg_key = cfg_key.decode('utf-8')
pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
# Get the scores and bounding boxes
scores = rpn_cls_prob[:, :, :, num_anchors:]
rpn_bbox_pred = rpn_bbox_pred.view((-1, 4))
scores = scores.contiguous().view(-1, 1)
proposals = bbox_transform_inv(anchors, rpn_bbox_pred)
proposals = clip_boxes(proposals, im_info[:2])
# Pick the top region proposals
scores, order = scores.view(-1).sort(descending=True)
if pre_nms_topN > 0:
order = order[:pre_nms_topN]
scores = scores[:pre_nms_topN].view(-1, 1)
proposals = proposals[order.data, :]
# Non-maximal suppression
keep = nms(proposals, scores.squeeze(1), nms_thresh)
# Pick th top region proposals after NMS
if post_nms_topN > 0:
keep = keep[:post_nms_topN]
proposals = proposals[keep, :]
scores = scores[keep, ]
# Only support single image as input
batch_inds = proposals.new_zeros(proposals.size(0), 1)
blob = torch.cat((batch_inds, proposals), 1)
return blob, scores
