def _add_proposals_from_file(
self, roidb, proposal_file, min_proposal_size, top_k, crowd_thresh
):
"""Add proposals from a proposals file to an roidb."""
logger.info('Loading proposals from: {}'.format(proposal_file))
with open(proposal_file, 'r') as f:
proposals = pickle.load(f)
id_field = 'indexes' if 'indexes' in proposals else 'ids' # compat fix
_sort_proposals(proposals, id_field)
box_list = []
for i, entry in enumerate(roidb):
if i % 2500 == 0:
logger.info(' {:d}/{:d}'.format(i + 1, len(roidb)))
boxes = proposals['boxes'][i]
# Sanity check that these boxes are for the correct image id
assert entry['id'] == proposals[id_field][i]
# Remove duplicate boxes and very small boxes and then take top k
boxes = box_utils.clip_boxes_to_image(
boxes, entry['height'], entry['width']
)
keep = box_utils.unique_boxes(boxes)
boxes = boxes[keep, :]
keep = box_utils.filter_small_boxes(boxes, min_proposal_size)
boxes = boxes[keep, :]
if top_k > 0:
boxes = boxes[:top_k, :]
box_list.append(boxes)
_merge_proposal_boxes_into_roidb(roidb, box_list)
if crowd_thresh > 0:
_filter_crowd_proposals(roidb, crowd_thresh)
def _add_proposals_from_file(self, roidb, proposal_file, min_proposal_size,
top_k, crowd_thresh):
"""Add proposals from a proposals file to an roidb."""
logger.info('Loading proposals from: {}'.format(proposal_file))
with open(proposal_file, 'r') as f:
proposals = pickle.load(f)
id_field = 'indexes' if 'indexes' in proposals else 'ids' # compat fix
_sort_proposals(proposals, id_field)
box_list = []
for i, entry in enumerate(roidb):
if i % 2500 == 0:
logger.info(' {:d}/{:d}'.format(i + 1, len(roidb)))
boxes = proposals['boxes'][i]
# Sanity check that these boxes are for the correct image id
assert entry['id'] == proposals[id_field][i]
# Remove duplicate boxes and very small boxes and then take top k
boxes = box_utils.clip_boxes_to_image(boxes, entry['height'],
entry['width'])
keep = box_utils.unique_boxes(boxes)
boxes = boxes[keep, :]
keep = box_utils.filter_small_boxes(boxes, min_proposal_size)
boxes = boxes[keep, :]
if top_k > 0:
boxes = boxes[:top_k, :]
box_list.append(boxes)
_merge_proposal_boxes_into_roidb(roidb, box_list)
if crowd_thresh > 0:
_filter_crowd_proposals(roidb, crowd_thresh)
def forward(self, inputs, outputs):
up_scale = self.up_scale
M = self.resolution
for lvl in range(self.num_fpn_lvls):
feat = inputs[2 * lvl].data
rois = inputs[2 * lvl + 1].data
num_rois = rois.shape[0]
spatial_scale = self.spatial_scales[lvl]
# convert from NCHW to NHWC
feat = feat.transpose((0, 2, 3, 1))
feat_h, feat_w = feat.shape[1], feat.shape[2]
# pad rois and narrow to the feature map scale
pad_rois = box_utils.expand_boxes(rois[:, 1:5], up_scale)
pad_rois = pad_rois * spatial_scale
pad_rois = box_utils.clip_boxes_to_image(pad_rois, feat_h, feat_w)
# abstact feature from the pad_rois
pad_roi_feats = np.zeros((num_rois, M, M, feat.shape[3]))
batch_idx = rois[:, 0]
for i in range(num_rois):
batch_id = int(batch_idx[i])
pad_roi = pad_rois[i]
pad_roi_feat = feat[batch_id, pad_roi[1]:pad_roi[3] + 1,
pad_roi[0]:pad_roi[2] + 1, :]
pad_roi_feat_resize = cv2.resize(pad_roi_feat, (M, M))
pad_roi_feats[i] = pad_roi_feat_resize
pad_roi_feats = pad_roi_feats.transpose((0, 3, 1, 2))
outputs[lvl].reshape(pad_roi_feats.shape)
outputs[lvl].data[...] = pad_roi_feats
def _add_proposals_from_file(self, roidb, proposal_file, min_proposal_size,
top_k):
"""Add proposals from a proposals file to an roidb.
"""
logger.info('Loading proposals from: {}'.format(proposal_file))
with open(proposal_file, 'r') as f:
proposals = pickle.load(f)
#proposals[seq_idx][idx]
box_list = []
for i, entry in enumerate(roidb):
if i % 500 == 0:
logger.info(' {:d}/{:d}'.format(i + 1, len(roidb)))
seq_idx = entry['seq_idx']
idx = entry['idx']
boxes = proposals['boxes'][seq_idx][idx]
# Remove duplicate boxes and very small boxes and then take top k
boxes = box_utils.clip_boxes_to_image(boxes, entry['height'],
entry['width'])
keep = box_utils.unique_boxes(boxes)
boxes = boxes[keep, :]
keep = box_utils.filter_small_boxes(boxes, min_proposal_size)
boxes = boxes[keep, :]
if top_k > 0:
boxes = boxes[:top_k, :]
box_list.append(boxes)
_merge_proposal_boxes_into_roidb(roidb, box_list)
def get_context_rois(self, rois, im_info, zoom_ratio=1.1):
"""Return the rois with more context.
- 'im_info': 2D tensor of shape (N, 3) where the three columns encode
the input image's [height, width, scale]. Height and width are
for the input to the network, not the original image; scale is the
scale factor used to scale the original image to the network input
size.
- 'rpn_rois': 2D tensor of shape (R, 5), for R RPN proposals where the
five columns encode [batch ind, x1, y1, x2, y2]. The boxes are
w.r.t. the network input, which is a *scaled* version of the
original image; these proposals must be scaled by 1 / scale (where
scale comes from im_info; see above) to transform it back to the
original input image coordinate system.
"""
if zoom_ratio > 0:
roi_boxes = rois[:, 1:]
batch_indice = rois[:, 0]
roi_boxes = boxes_utils.expand_boxes(roi_boxes, zoom_ratio)
roi_boxes = boxes_utils.clip_boxes_to_image(roi_boxes, im_info[0], im_info[1])
rois = np.concatenate((batch_indice, roi_boxes), axis=1)
else:
# Add global context, i.e. roi-pooling on the whole feature map.
rois[:, 1] = 0
rois[:, 2] = 0
rois[:, 3] = 800 - 1
rois[:, 4] = 800 - 1
#rois[:, 3] = im_info[0][1] - 1
#rois[:, 4] = im_info[0][0] - 1
return rois
def forward(self, inputs, outputs):
data = inputs[0].data
k_min = self.k_min
k_max = self.k_max
up_scale = self.up_scale
height, width = data.shape[2], data.shape[3]
for lvl in range(k_min, k_max + 1):
rois = inputs[1 + lvl - k_min].data # skip the 'data' blob
bboxes = rois[:, 1:5]
batch_ids = rois[:, [0]]
# up-scale the bboxes and narrow down to image boundary
pad_bboxes = box_utils.expand_boxes(bboxes, up_scale)
pad_bboxes = box_utils.clip_boxes_to_image(pad_bboxes, height,
width)
# add the batch_ids to the rois
pad_rois = np.hstack((batch_ids, pad_bboxes))
outputs[lvl - k_min].reshape(pad_rois.shape)
outputs[lvl - k_min].data[...] = pad_rois
# copy rois_idx_restore_int32 to the scale_rois_idx_restore_int32
# A little surgery for int32 type requirement
rois_idx_restore_int32 = inputs[-1].data
outputs[-1].init(list(rois_idx_restore_int32.shape),
caffe2_pb2.TensorProto.INT32)
outputs[-1].reshape(rois_idx_restore_int32.shape)
outputs[-1].data[...] = rois_idx_restore_int32.astype(np.int32)
def add_refine_keypoints_blobs_gaussian(blobs, roidb, fg_rois_per_image,
fg_inds, im_scale, batch_idx, data):
"""Add Mask R-CNN keypoint specific blobs to the given blobs dictionary."""
# Note: gt_inds must match how they're computed in
# datasets.json_dataset._merge_proposal_boxes_into_roidb
gt_inds = np.where(roidb['gt_classes'] > 0)[0]
gt_keypoints = roidb['gt_keypoints']
# Load the kp_fg_inds generated by keypoint_rcnn.py. So we avoid the issue
# of mismatched keypoint_rois and refined_keypoint_rois, which cause a big
# issue for training.
kp_fg_inds = blobs['keypoint_fg_inds']
if kp_fg_inds.shape[0] > 0:
sampled_fg_rois = roidb['boxes'][kp_fg_inds]
box_to_gt_ind_map = roidb['box_to_gt_ind_map'][kp_fg_inds]
# Let's expand the rois
up_scale = cfg.REFINENET.UP_SCALE
inp_h, inp_w = data.shape[2], data.shape[3]
pad_img_h, pad_img_w = inp_h / im_scale, inp_w / im_scale
pad_fg_rois = box_utils.expand_boxes(sampled_fg_rois, up_scale)
pad_fg_rois = box_utils.clip_boxes_to_image(pad_fg_rois, pad_img_h,
pad_img_w)
num_keypoints = gt_keypoints.shape[2]
sampled_keypoints = -np.ones(
(len(pad_fg_rois), gt_keypoints.shape[1], num_keypoints),
dtype=gt_keypoints.dtype)
for ii in range(len(pad_fg_rois)):
ind = box_to_gt_ind_map[ii]
if ind >= 0:
sampled_keypoints[ii, :, :] = gt_keypoints[gt_inds[ind], :, :]
assert np.sum(sampled_keypoints[ii, 2, :]) > 0
heats, weights = keypoint_utils.keypoints_to_gaussian_heatmap_labels(
sampled_keypoints, pad_fg_rois, M=cfg.REFINENET.KRCNN.HEATMAP_SIZE)
else: # If there are no fg keypoint rois (it does happen)
# The network cannot handle empty blobs, so we must provide a heatmap
# We simply take the first bg roi, given it an all zero heatmap, and
# set its weights to zero (ignore label).
roi_inds = np.where(roidb['gt_classes'] == 0)[0]
# sampled_fg_rois is actually one random roi, but that's ok because ...
pad_fg_rois = roidb['boxes'][roi_inds[0]].reshape((1, -1))
# We give it an 0's blob
M = cfg.REFINENET.KRCNN.HEATMAP_SIZE
heats = blob_utils.zeros((1, cfg.KRCNN.NUM_KEYPOINTS, M, M))
# We set weights to 0 (ignore label)
weights = blob_utils.zeros((1, cfg.KRCNN.NUM_KEYPOINTS, 1))
pad_fg_rois *= im_scale
repeated_batch_idx = batch_idx * blob_utils.ones((pad_fg_rois.shape[0], 1))
pad_fg_rois = np.hstack((repeated_batch_idx, pad_fg_rois))
blobs['refined_keypoint_rois'] = pad_fg_rois
blobs['refined_keypoint_heatmaps'] = heats
blobs['refined_keypoint_weights'] = weights
def forward(self, inputs, outputs):
data = inputs[0].data
mask_probs = inputs[1].data
mask_rois = inputs[2].data
# whether using binary threshold for indicator
if cfg.REFINENET.USE_PERCENTTHRES_INDICATOR:
mask_probs_reshape = mask_probs.reshape(mask_probs.shape[0],
mask_probs.shape[1],-1)
thres = int(cfg.REFINENET.PERCENTINDICATOR_THRES * \
mask_probs_reshape.shape[2])
mask_probs_sort = np.argsort(-mask_probs_reshape, axis=2)[:,:,:thres]
mask_binary = np.zeros(mask_probs_reshape.shape, dtype=np.float32)
for i in range(mask_probs_sort.shape[0]):
for j in range(mask_probs_sort.shape[1]):
mask_binary[i,j,mask_probs_sort[i,j]] = 1.
mask_binary = mask_binary.reshape(mask_probs.shape)
mask_probs *= mask_binary
elif cfg.REFINENET.USE_THRES_INDICATOR:
mask_binary = np.array(
mask_probs > cfg.REFINENET.INDICATOR_THRES, dtype=np.float32
)
mask_probs *= mask_binary
# output indicator resolution
M = self.resolution
up_scale = self.up_scale
num_cls = mask_probs.shape[1]
num_rois = mask_rois.shape[0]
mask_indicators = np.zeros((num_rois, M, M, num_cls), dtype=np.float32)
# preparing data
height, width = data.shape[2], data.shape[3]
mask_probs_NHWC = mask_probs.transpose((0,2,3,1))
rois = mask_rois[:, 1:5] # ignore batch_id
pad_rois = box_utils.expand_boxes(rois, up_scale)
pad_rois = box_utils.clip_boxes_to_image(pad_rois, height, width)
# calculate converted coordinates
converted_coords = box_utils.convert_coordinate(rois, pad_rois, M)
for i in range(num_rois):
mask_prob = mask_probs_NHWC[i]
coords = converted_coords[i]
shape = (coords[2]-coords[0]+1, coords[3]-coords[1]+1) # w,h
if shape[0] < 1 or shape[1] < 1:
continue
mask_prob_resize = cv2.resize(mask_prob, shape)
if mask_prob_resize.shape[2] == 1:
mask_prob_resize = mask_prob_resize[:, :, np.newaxis]
mask_indicators[i, coords[1]:coords[3]+1, coords[0]:coords[2]+1] = \
mask_prob_resize
swap_order = (0, 3, 1, 2)
mask_indicators = mask_indicators.transpose(swap_order)
outputs[0].reshape(mask_indicators.shape)
outputs[0].data[...] = mask_indicators
def _add_proposals_from_file(self, roidb, proposal_file, min_proposal_size,
top_k, crowd_thresh):
"""Add proposals from a proposals file to an roidb."""
logger.info('Loading proposals from: {}'.format(proposal_file))
# with open(proposal_file, 'rb') as f:
# proposals = pickle.load(f, encoding='bytes')
with open(proposal_file, 'rb') as f:
proposals1 = pickle.load(f, encoding='bytes')
#[b'cfg', b'boxes', b'scores', b'ids', b'names']
proposals = {}
proposals['cfg'] = proposals1[b'cfg']
proposals['boxes'] = proposals1[b'boxes']
proposals['scores'] = proposals1[b'scores']
proposals['ids'] = proposals1[b'ids']
proposals['names'] = proposals1[b'names']
#proposals = pd.read_pickle(proposal_file)
print(proposals.keys())
print(
np.array(proposals['boxes']).shape,
"!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
# proposals = self.convert(proposals)
print(proposals.keys())
id_field = 'indexes' if 'indexes' in proposals else 'ids' # compat fix
_sort_proposals(proposals, id_field)
box_list = []
for i, entry in enumerate(roidb):
if i % 2500 == 0:
logger.info(' {:d}/{:d}'.format(i + 1, len(roidb)))
boxes = proposals['boxes'][i]
print("entry id in entry['id'] ", entry['id'])
#print(proposals[id_field][i])
# Sanity check that these boxes are for the correct image id
assert entry['id'] == proposals[id_field][i]
# Remove duplicate boxes and very small boxes and then take top k
boxes = box_utils.clip_boxes_to_image(boxes, entry['height'],
entry['width'])
# keep = box_utils.unique_boxes(boxes)
# boxes = boxes[keep, :]
# keep = box_utils.filter_small_boxes(boxes, min_proposal_size)
# boxes = boxes[keep, :]
# if top_k > 0:
# boxes = boxes[:top_k, :]
box_list.append(boxes)
print("shape box_list!!!!!!!!!!!!!!!!!!!!!!!!! ", np.array(box_list))
_merge_proposal_boxes_into_roidb(roidb, box_list)
if crowd_thresh > 0:
_filter_crowd_proposals(roidb, crowd_thresh)
def forward(self, inputs, outputs):
data = inputs[0].data
rois = inputs[1].data
up_scale = self.up_scale
height, width = data.shape[2], data.shape[3]
bboxes = rois[:, 1:5]
batch_ids = rois[:, [0]]
# up-scale the bboxes and clip to image boundary
# pad bboxes and narrow to the feature map scale
pad_bboxes = box_utils.expand_boxes(bboxes, up_scale)
pad_bboxes = box_utils.clip_boxes_to_image(pad_bboxes, height, width)
# add the batch_ids to the rois
pad_rois = np.hstack((batch_ids, pad_bboxes))
outputs[0].reshape(pad_rois.shape)
outputs[0].data[...] = pad_rois
def forward(self, inputs, outputs):
data = inputs[0].data
keypoint_probs = inputs[1].data
keypoint_rois = inputs[2].data
# output indicator resolution
M = self.resolution
up_scale = self.up_scale
num_rois = keypoint_rois.shape[0]
num_keypoints = keypoint_probs.shape[1]
# first expand the keypoint rois
height, width = data.shape[2], data.shape[3]
pad_rois = box_utils.expand_boxes(keypoint_rois[:, 1:5], up_scale)
pad_rois = box_utils.clip_boxes_to_image(pad_rois, height, width)
# get keypoint predictions and their probs
# output shape is (#rois, 3, #keypoints) and 3 means (x, y, prob)
pred_rois = keypoint_utils.probs_to_keypoints(keypoint_probs, keypoint_rois)
# map keypoint position to the pad_rois
# output shape is (#rois, #keypoints), locations flatter out
locations_on_pad_rois, _ = keypoint_utils.keypoints_to_heatmap_labels(
pred_rois, pad_rois, M
)
locations_on_pad_rois = locations_on_pad_rois.astype(np.int32)
# and now generate keypoint indicators
keypoint_indicators = blob_utils.zeros((num_rois, num_keypoints, M**2))
for i in range(num_rois):
locations = locations_on_pad_rois[i] # shape (#keypoints, )
for k in range(num_keypoints):
keypoint_indicators[i, k, locations[k]] = pred_rois[i, 2, k]
# and reshape to 4 dimension
keypoint_indicators = keypoint_indicators.reshape(
(num_rois, num_keypoints, M, M)
)
outputs[0].reshape(keypoint_indicators.shape)
outputs[0].data[...] = keypoint_indicators
def add_refine_local_mask_blobs(blobs, sampled_boxes, roidb, im_scale,
batch_idx, data):
"""Add RefineNet Mask specific blobs to the input blob dictionary."""
# Prepare the mask targets by associating one gt mask to each training roi
# that has a fg (non-bg) class label.
M = cfg.REFINENET.RESOLUTION
up_scale = cfg.REFINENET.UP_SCALE
polys_gt_inds = np.where((roidb['gt_classes'] > 0)
& (roidb['is_crowd'] == 0))[0]
gt_classes = roidb['gt_classes'][polys_gt_inds]
polys_gt = [roidb['segms'][i] for i in polys_gt_inds]
boxes_from_polys = segm_utils.polys_to_boxes(polys_gt)
fg_inds = np.where(blobs['labels_int32'] > 0)[0]
roi_has_mask = blobs['labels_int32'].copy()
roi_has_mask[roi_has_mask > 0] = 1
# Define size variables
inp_h, inp_w = data.shape[2], data.shape[3]
pad_img_h, pad_img_w = inp_h / im_scale, inp_w / im_scale
if fg_inds.shape[0] > 0:
# Class labels for the foreground rois
mask_class_labels = blobs['labels_int32'][fg_inds]
masks = blob_utils.zeros((fg_inds.shape[0], M**2), int32=True)
# Find overlap between all foreground rois and the bounding boxes
# enclosing each segmentation
rois_fg = sampled_boxes[fg_inds]
overlaps_bbfg_bbpolys = box_utils.bbox_overlaps(
rois_fg.astype(np.float32, copy=False),
boxes_from_polys.astype(np.float32, copy=False))
# Map from each fg rois to the index of the mask with highest overlap
# (measured by bbox overlap)
fg_polys_inds = np.argmax(overlaps_bbfg_bbpolys, axis=1)
# Expand the foreground rois by a factor of up_scale and
# clip by the padded image boundary
pad_rois_fg = box_utils.expand_boxes(rois_fg, up_scale)
pad_rois_fg = box_utils.clip_boxes_to_image(pad_rois_fg, pad_img_h,
pad_img_w)
if cfg.REFINENET.ONLY_USE_CROWDED_SAMPLES:
# Only use crowded samples to train the RefineNet
THRES = cfg.REFINENET.OVERLAP_THRESHOLD
for i in range(rois_fg.shape[0]):
overlap = overlaps_bbfg_bbpolys[i]
if np.sum(overlap > THRES) > 1:
# if has multiple instances overlapped, use it for training
fg_polys_ind = fg_polys_inds[i]
poly_gt = polys_gt[fg_polys_ind]
pad_roi_fg = pad_rois_fg[i]
# Rasterize the portion of the polygon mask within the given fg roi
# to an M x M binary image
mask = segm_utils.polys_to_mask_wrt_box(
poly_gt, pad_roi_fg, M)
mask = np.array(mask > 0,
dtype=np.int32) # Ensure it's binary
masks[i, :] = np.reshape(mask, M**2)
else: # Only one instance, then set label to be -1 (ignored)
masks[i, :] = -1
mask_class_labels[i] = 0
elif cfg.REFINENET.ASSIGN_LARGER_WEIGHT_FOR_CROWDED_SAMPLES:
loss_weights = blob_utils.ones((rois_fg.shape[0], ))
for i in range(rois_fg.shape[0]):
fg_polys_ind = fg_polys_inds[i]
poly_gt = polys_gt[fg_polys_ind]
pad_roi_fg = pad_rois_fg[i]
class_label = mask_class_labels[i]
# Rasterize the portion of the polygon mask within the given
# fg roi to an M x M binary image
mask = segm_utils.polys_to_mask_wrt_box(poly_gt, pad_roi_fg, M)
mask = np.array(mask > 0, dtype=np.int32) # Ensure it's binary
masks[i, :] = np.reshape(mask, M**2)
# And now determine the weight for each roi. If any instance
# that is of the same class as the RoI, then we expect it to
# be a hard sample and assigns a larger weight for this RoI
for j in range(len(polys_gt)):
if j == fg_polys_ind:
continue
if gt_classes[
j] == class_label: # only same class is valid
mask = segm_utils.polys_to_mask_wrt_box(
polys_gt[j], pad_roi_fg, M)
# and check if has anypart fall inside the bbox
is_inside_bbox = (np.sum(mask) > 0)
if is_inside_bbox:
loss_weights[i] = cfg.REFINENET.WEIGHT_LOSS_CROWDED
break # early stop
else:
# add fg targets
for i in range(rois_fg.shape[0]):
fg_polys_ind = fg_polys_inds[i]
poly_gt = polys_gt[fg_polys_ind]
pad_roi_fg = pad_rois_fg[i]
# Rasterize the portion of the polygon mask within the given fg roi
# to an M x M binary image
mask = segm_utils.polys_to_mask_wrt_box(poly_gt, pad_roi_fg, M)
mask = np.array(mask > 0, dtype=np.int32) # Ensure it's binary
masks[i, :] = np.reshape(mask, M**2)
else: # If there are no fg masks (it does happen)
# The network cannot handle empty blobs, so we must provide a mask
# We simply take the first bg roi, given it an all -1's mask (ignore
# label), and label it with class zero (bg).
bg_inds = np.where(blobs['labels_int32'] == 0)[0]
# pad_rois_fg is actually one background roi, but that's ok because ...
pad_rois_fg = sampled_boxes[bg_inds[0]].reshape((1, -1))
# We give it an -1's blob (ignore label)
masks = -blob_utils.ones((1, M**2), int32=True)
# We label it with class = 0 (background)
mask_class_labels = blob_utils.zeros((1, ))
# Mark that the first roi has a mask
roi_has_mask[0] = 1
if cfg.MRCNN.CLS_SPECIFIC_MASK:
masks = _expand_to_class_specific_mask_targets(masks,
mask_class_labels)
# Scale rois_fg and format as (batch_idx, x1, y1, x2, y2)
pad_rois_fg = (pad_rois_fg.astype(np.float32)) * im_scale
repeated_batch_idx = batch_idx * blob_utils.ones((pad_rois_fg.shape[0], 1))
pad_rois_fg = np.hstack((repeated_batch_idx, pad_rois_fg)).astype(np.int32)
# Update blobs dict with Refine-Net blobs
blobs['refined_mask_rois'] = pad_rois_fg
blobs['roi_has_refined_mask_int32'] = roi_has_mask
blobs['refined_masks_int32'] = masks
if cfg.REFINENET.ASSIGN_LARGER_WEIGHT_FOR_CROWDED_SAMPLES:
blobs['loss_weights'] = loss_weights
def __init__(self,
root,
year='2007',
image_set='trainval',
download=False,
transforms=None):
self.root = root
self.transforms = transforms
self.year = year
if year == "2007" and image_set == "test":
year = "2007-test"
self.url = DATASET_YEAR_DICT[year]['url']
self.filename = DATASET_YEAR_DICT[year]['filename']
self.md5 = DATASET_YEAR_DICT[year]['md5']
valid_sets = ["train", "trainval", "val"]
if year == "2007-test":
valid_sets.append("test")
self.image_set = verify_str_arg(image_set, "image_set", valid_sets)
base_dir = DATASET_YEAR_DICT[year]['base_dir']
voc_root = os.path.join(self.root, base_dir)
image_dir = os.path.join(voc_root, 'JPEGImages')
annotation_dir = os.path.join(voc_root, 'Annotations')
if download:
download_extract(self.url, self.root, self.filename, self.md5)
# pause()
if not os.path.isdir(voc_root):
raise RuntimeError('Dataset not found or corrupted.' +
' You can use download=True to download it')
splits_dir = os.path.join(voc_root, 'ImageSets/Main')
split_f = os.path.join(splits_dir, image_set.rstrip('\n') + '.txt')
with open(os.path.join(split_f), "r") as f:
file_names = [x.strip() for x in f.readlines()]
self.class_labels = ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor']
self.num_classes = len(self.class_labels)
self.images = [os.path.join(image_dir, x + ".jpg") for x in file_names]
annotations = [os.path.join(annotation_dir, x + ".xml") for x in file_names]
assert (len(self.images) == len(annotations))
#load annotations:
self.annotations = []
for i, ann_path in enumerate(annotations):
raw_annot = self.parse_voc_xml(ET.parse(ann_path).getroot())['annotation']
self.annotations.append({
'img_key': int(raw_annot['filename'].replace('.jpg','')),
'width' : int(raw_annot['size']['width']),
'height' : int(raw_annot['size']['height']),
'object' : raw_annot['object']
})
##TODO ajust this to be more beauty =p
ss_data = self.root + 'selective_search_data/voc_' + self.year + '_' + self.image_set + '.pkl'
with open(ss_data, 'rb') as f:
proposals = pickle.load(f)
sort_proposals(proposals, 'indexes')
self.proposals = []
for i, boxes in enumerate(proposals['boxes']):
if i % 2500 == 0:
logger.info(' {:d}/{:d}'.format(i + 1, len(proposals['boxes'])))
annotation = self.annotations[i]
assert annotation['img_key'] == proposals['indexes'][i]
# Remove duplicate boxes and very small boxes and then take top k
boxes = box_utils.clip_boxes_to_image(boxes, annotation['height'], annotation['width'])
keep = box_utils.unique_boxes(boxes)
boxes = boxes[keep, :]
keep = box_utils.filter_small_boxes(boxes, cfg.FAST_RCNN.MIN_PROPOSAL_SIZE)
boxes = boxes[keep, :]
if cfg.FAST_RCNN.TOP_K > 0:
boxes = boxes[:cfg.FAST_RCNN.TOP_K, :]
self.proposals.append(boxes.astype(np.float))
