def _load_coco_annotation(self, index):
"""
Loads COCO bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
im_ann = self._COCO.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
annIds = self._COCO.getAnnIds(imgIds=index, iscrowd=None)
objs = self._COCO.loadAnns(annIds)
# Sanitize bboxes -- some are invalid
valid_objs = []
for obj in objs:
x1 = np.max((0, obj['bbox'][0]))
y1 = np.max((0, obj['bbox'][1]))
x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
valid_objs.append(obj)
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
# Lookup table to map from COCO category ids to our internal class
# indices
coco_cat_id_to_class_ind = dict([(self._class_to_coco_cat_id[cls],
self._class_to_ind[cls])
for cls in self._classes[1:]])
for ix, obj in enumerate(objs):
cls = coco_cat_id_to_class_ind[obj['category_id']]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
seg_areas[ix] = obj['area']
if obj['iscrowd']:
# Set overlap to -1 for all classes for crowd objects
# so they will be excluded during training
overlaps[ix, :] = -1.0
else:
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False,
'seg_areas': seg_areas
}
def _load_coco_annotation(self, index):
"""
Loads COCO bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
im_ann = self._COCO.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
annIds = self._COCO.getAnnIds(imgIds=index, iscrowd=None)
objs = self._COCO.loadAnns(annIds)
# Sanitize bboxes -- some are invalid
valid_objs = []
for obj in objs:
x1 = np.max((0, obj['bbox'][0]))
y1 = np.max((0, obj['bbox'][1]))
x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
valid_objs.append(obj)
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
# Lookup table to map from COCO category ids to our internal class
# indices
coco_cat_id_to_class_ind = dict([(self._class_to_coco_cat_id[cls],
self._class_to_ind[cls])
for cls in self._classes[1:]])
for ix, obj in enumerate(objs):
cls = coco_cat_id_to_class_ind[obj['category_id']]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
seg_areas[ix] = obj['area']
if obj['iscrowd']:
# Set overlap to -1 for all classes for crowd objects
# so they will be excluded during training
overlaps[ix, :] = -1.0
else:
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {'boxes' : boxes,
'gt_classes': gt_classes,
'gt_overlaps' : overlaps,
'flipped' : False,
'seg_areas' : seg_areas}
def _load_display_annotation(self, index):
"""
Loads Display bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
annotation_file = os.path.join(
self._data_path,
'annotations/instances_{}.txt'.format(self._image_set))
assert os.path.isfile(annotation_file), annotation_file
txt_annotations = open(annotation_file, 'r')
annotations = txt_annotations.readlines()
num_objs = 3
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
width, height = 0, 0
for i in range(0, len(annotations), 3):
if i != index:
continue
temp = annotations[i].split(',')
# data in ground truth file has 3 line for each img
for j in range(0, 3):
temp = annotations[i + j].split(',')
x1 = int(temp[1])
y1 = int(temp[2])
x2 = int(temp[3])
y2 = int(temp[4])
width = x2 - x1
height = y2 - y1
box = [x1, y1, x2, y2]
boxes[j, :] = box
gt_classes[j] = int(temp[5][0]) + 1
seg_areas[j] = (x2 - x1) * (y2 - y1)
overlaps[j, int(temp[5][0]) + 1] = 1.0
print("===================[display.py:173] ", overlaps)
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
print("===================[display.py:173] ", overlaps)
return {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False,
'seg_areas': seg_areas
}
def _loadCarlaAnnotation(self, index):
"""
Load the annotations of the ith image from the info list
"""
width, height = 1920, 1080
objList = np.array(self.infos[self.image_id_at(index)])
nObjects = len(objList)
# Ignores
ignores = np.array([False] * nObjects, dtype = int)
# Ground-truth class
objClasses = ds.getClasses(objList, self.classes[0]) # Class name of objects
gtClassses = np.array([self.cls2Ind[c] for c in objClasses]) # convert to the index of class
# Boxes
boxes = ds.get2dBoxes(objList)
# 3d object centers
locations = ds.get3dLocations(objList)
centers = self._convertTo(locations)
# Overlaps (one-hot vector form of labels)
overlaps = np.zeros((nObjects, self.num_classes), dtype = float)
overlaps[np.arange(nObjects), gtClassses] += 1
overlaps = scipy.sparse.csr_matrix(overlaps) # convert to a sparse matrix
# Area of segments
segAreas = self._computeSegAreas(boxes)
# End-of-video id (flag)
endvids = np.zeros(nObjects, dtype = int)
if (index == len(self.infos) - 1):
# Last image of the video, set them to 1
endvids += 1
# Validation boxes
ds.validate_boxes(boxes, width=width, height=height)
info_set = {
"width": width,
"height": height,
"boxes": boxes,
"gt_classes": gtClassses,
"gt_overlaps": overlaps,
"flipped": False,
"seg_areas": segAreas,
"ignore": ignores,
"end_vid": endvids,
"center": centers
}
return info_set
def _load_proposals(self, method, gt_roidb):
"""
Load pre-computed proposals in the format provided by Jan Hosang:
http://www.mpi-inf.mpg.de/departments/computer-vision-and-multimodal-
computing/research/object-recognition-and-scene-understanding/how-
good-are-detection-proposals-really/
For MCG, use boxes from http://www.eecs.berkeley.edu/Research/Projects/
CS/vision/grouping/mcg/ and convert the file layout using
lib/datasets/tools/mcg_munge.py.
"""
box_list = []
imsize_list = []
top_k = self.config['top_k']
valid_methods = [
'MCG',
'selective_search',
'edge_boxes_AR',
'edge_boxes_70']
assert method in valid_methods
print 'Loading {} boxes'.format(method)
for i, index in enumerate(self._image_index):
if i % 1000 == 0:
print '{:d} / {:d}'.format(i + 1, len(self._image_index))
box_file = osp.join(
cfg.DATA_DIR, 'coco_proposals', method, 'mat',
self._get_box_file(index))
raw_data = sio.loadmat(box_file)['boxes']
boxes = np.maximum(raw_data - 1, 0).astype(np.uint16)
if method == 'MCG':
# Boxes from the MCG website are in (y1, x1, y2, x2) order
boxes = boxes[:, (1, 0, 3, 2)]
im_ann = self._COCO.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
imsize_list.append([height,width])
# Remove duplicate boxes and very small boxes and then take top k
keep = ds_utils.unique_boxes(boxes)
boxes = boxes[keep, :]
keep = ds_utils.filter_small_boxes(boxes, self.config['min_size'])
boxes = boxes[keep, :]
boxes = boxes[:top_k, :]
boxes[:,0]=np.maximum(boxes[:,0]-cfg.MARGIN,0)
boxes[:,2]=np.minimum(boxes[:,2]+cfg.MARGIN,width-1)
boxes[:,1]=np.maximum(boxes[:,1]-cfg.MARGIN,0)
boxes[:,3]=np.minimum(boxes[:,3]+cfg.MARGIN,height-1)
box_list.append(boxes)
# Sanity check
ds_utils.validate_boxes(boxes, width=width, height=height)
return self.create_roidb_from_box_list(box_list, gt_roidb,imsize_list)
def _load_vg_annotation(self, index):
"""
Loads VG bounding-box instance annotations.
"""
img = self._VG.load_imgs(index)[0]
width = img['width']
height = img['height']
ann_ids = self._VG.get_ann_ids(img_ids=index)
anns = self._VG.load_anns(ann_ids)
# Sanitize bboxes -- some are invalid
valid_anns = []
for ann in anns:
x1 = np.max((0, ann['x']))
y1 = np.max((0, ann['y']))
x2 = np.min((width - 1, x1 + np.max((0, ann['w'] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, ann['h'] - 1))))
if x2 >= x1 and y2 >= y1:
ann['clean_bbox'] = [x1, y1, x2, y2]
valid_anns.append(ann)
anns = valid_anns
num_anns = len(anns)
boxes = np.zeros((num_anns, 4), dtype=np.uint16)
gt_classes = np.zeros((num_anns), dtype=np.int32)
overlaps = np.zeros((num_anns, self.num_classes), dtype=np.float32)
# # Lookup table to map from VG category ids to our internal class
# # indices
# vg_id_to_class_ind = dict([(self._class_to_vg_id[cls], self._class_to_ind[cls])
# for cls in self._classes[1:]])
for ix, ann in enumerate(anns):
cls = self._vg_id_to_class_ind[ann['category_id']]
boxes[ix, :] = ann['clean_bbox']
gt_classes[ix] = cls
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False
}
def _load_vg_annotation(self, index):
"""
Loads visual genome bounding-box instance annotations.
"""
object_anno = self._VG.loadImgsAnno(index)
height, width = self._VG.getImageWidthHeights(index)
objs = object_anno['objects']
# Sanitize bboxes -- some are invalid
valid_objs = []
for obj in objs:
if obj['names'][0] in self.cats:
x1 = np.max((0, obj['x']))
y1 = np.max((0, obj['y']))
x2 = np.min((width - 1, x1 + np.max((0, obj['w'] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['h'] - 1))))
obj['area'] = obj['w'] * obj['h']
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
valid_objs.append(obj)
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
for ix, obj in enumerate(objs):
cls = self._class_to_ind[obj["names"][0]]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
seg_areas[ix] = obj['area']
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
#print('processing time:{}'.format(time.time()-t))
return {'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False,
'seg_areas': seg_areas}
def _load_stacked_annotation(self, index, gtfile):
"""
Loads bounding-box instance annotations.
"""
gtboxes, classes, gtids = self.idx2gtinstance(index, gtfile)
width = self._widths[index]
height = self._heights[index]
num_objs = len(gtboxes)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes),
dtype=np.float32) # 1 at the gt class, 0 otherwise
seg_areas = np.zeros((num_objs), dtype=np.float32) # box area here
gt_ids = np.array(gtids, dtype=int)
for i, ibox in enumerate(gtboxes):
bbox = ibox - 1 # start at 0
x1 = bbox[0]
y1 = bbox[1]
x2 = bbox[2]
y2 = bbox[3]
boxes[i, :] = bbox
cls = classes[i]
gt_classes[i] = cls
overlaps[i, cls] = 1
seg_areas[i] = (x2 - x1 + 1) * (y2 - y1 + 1)
overlaps = scipy.sparse.csr_matrix(overlaps)
ds_utils.validate_boxes(boxes, width=width, height=height)
return {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'gt_ids': gt_ids,
'flipped': False,
'seg_areas': seg_areas
}
def _load_coco_text_annotation(self, index):
"""
coco returns dict with keys ('boxes', gt_classes',
'gt_overlaps', 'flipped', 'seg_areas')
"""
im_ann = self._ct.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
objs = self._ct.loadAnns(self._ct.getAnnIds(imgIds=index))
valid_objs = []
for obj in objs:
x1 = np.max((0, obj['bbox'][0]))
y1 = np.max((0, obj['bbox'][1]))
x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
valid_objs.append(obj)
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.ones((num_objs), dtype=np.int32) # always the same
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
for ix, obj in enumerate(objs):
boxes[ix, :] = obj['clean_bbox']
seg_areas[ix] = obj['area']
overlaps[ix, 1] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False,
'seg_areas': seg_areas
}
def _load_mot_annotation(self):
gt_roidb = []
annotationfile = os.path.join(self._data_path, 'gt_bbox.txt')
f = open(annotationfile)
split_line = f.readline().strip().split()
num = 1
while (split_line):
num_objs = int(split_line[1])
boxes = np.zeros((num_objs, 4), dtype=np.int32)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
for i in range(num_objs):
x1 = max(0, float(split_line[2 + i * 4]))
y1 = max(0, float(split_line[3 + i * 4]))
x2 = min(1919, float(split_line[4 + i * 4]))
y2 = min(1079, float(split_line[5 + i * 4]))
# x1 = float(split_line[2 + i * 4])
# y1 = float(split_line[3 + i * 4])
# x2 = float(split_line[4 + i * 4])
# y2 = float(split_line[5 + i * 4])
cls = self._class_to_ind['person']
boxes[i, :] = [x1, y1, x2, y2]
gt_classes[i] = cls
overlaps[i, cls] = 1.0
ds_utils.validate_boxes(boxes, 1920, 1080)
overlaps = scipy.sparse.csr_matrix(overlaps)
gt_roidb.append({
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False
})
split_line = f.readline().strip().split()
f.close()
return gt_roidb
def _load_coco_annotation(self, index):
"""
Loads COCO bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
if cfg.DEBUG:
print '_load_coco_annotation'
im_ann = self._COCO.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
annIds = self._COCO.getAnnIds(imgIds=index, iscrowd=None)
objs = self._COCO.loadAnns(annIds)
# Sanitize bboxes -- some are invalid
valid_objs = []
for obj in objs:
x1 = np.max((0, obj['bbox'][0]))
y1 = np.max((0, obj['bbox'][1]))
x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
valid_objs.append(obj)
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
masks = np.zeros((num_objs, height, width), dtype=np.uint16)
# Lookup table to map from COCO category ids to our internal class
# indices
coco_cat_id_to_class_ind = dict([(self._class_to_coco_cat_id[cls],
self._class_to_ind[cls])
for cls in self._classes[1:]])
for ix, obj in enumerate(objs):
cls = coco_cat_id_to_class_ind[obj['category_id']]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
seg_areas[ix] = obj['area']
if obj['iscrowd']:
# Set overlap to -1 for all classes for crowd objects
# so they will be excluded during training
overlaps[ix, :] = -1.0
else:
overlaps[ix, cls] = 1.0
#*************************************
# read mask_name for every instance
#*************************************
# path e.g., /data/coco/gt_mask/train2014/groundt_train_000000000009.png
mask_path = osp.join(self._data_path, 'gt_mask', self._data_name, obj['mask_name'])
mask = np.asarray(mat4py.loadmat(mask_path)['mask_gt'])
# mask_flat = np.asarray(list(PIL.Image.open(mask_path).getdata()))
# mask = mask_flat.reshape((height, width))
mask[np.where(mask != obj['id'])] = 0
mask[np.where(mask == obj['id'])] = 1
masks[ix,:,:] = mask
if cfg.DEBUG:
print masks.shape
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {'boxes' : boxes,
'gt_classes': gt_classes,
'gt_overlaps' : overlaps,
'flipped' : False,
'seg_areas' : seg_areas,
'masks' : masks}
def _load_coco_annotation(self, index):
"""
Loads COCO bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
im_ann = self._COCO.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
annIds = self._COCO.getAnnIds(imgIds=index, iscrowd=None)
objs = self._COCO.loadAnns(annIds)
# Sanitize bboxes -- some are invalid
valid_objs = []
count = 0
for obj in objs:
x1 = np.max((0, obj['bbox'][0]))
y1 = np.max((0, obj['bbox'][1]))
x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
if cfg.TRAIN.MASK_REG:
count = count + 1
obj['seg_mask_ind'] = [index, count]
seg_mask = self._COCO.annToMask(obj)
seg_mask_save = seg_mask
seg_mask_save = seg_mask_save.astype(bool)
seg_mask_path = './data/cache/' + 'GTsegmask_' + cfg.TRAIN.TRAINING_DATA + '/' + str(index) + '_' + str(count) + '_segmask.sm'
with open(seg_mask_path, 'wb') as f_seg_save:
cPickle.dump(seg_mask_save, f_seg_save, cPickle.HIGHEST_PROTOCOL)
valid_objs.append(obj)
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
seg_mask_inds = np.zeros((num_objs, 2), dtype=np.uint32)
# Lookup table to map from COCO category ids to our internal class
# indices
coco_cat_id_to_class_ind = dict([(self._class_to_coco_cat_id[cls],
self._class_to_ind[cls])
for cls in self._classes[1:]])
for ix, obj in enumerate(objs):
cls = coco_cat_id_to_class_ind[obj['category_id']]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
seg_areas[ix] = obj['area']
if cfg.TRAIN.MASK_REG:
seg_mask_inds[ix, :] = obj['seg_mask_ind']
if obj['iscrowd']:
# Set overlap to -1 for all classes for crowd objects
# so they will be excluded during training
overlaps[ix, :] = -1.0
else:
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
if not cfg.TRAIN.MASK_REG:
return {'boxes' : boxes,
'gt_classes' : gt_classes,
'gt_overlaps' : overlaps,
'flipped' : False,
'seg_areas' : seg_areas}
else:
return {'boxes' : boxes,
'gt_classes' : gt_classes,
'gt_overlaps' : overlaps,
'flipped' : False,
'seg_areas' : seg_areas,
'seg_mask_inds' : seg_mask_inds}
def _load_gta_annotation(self, index):
"""
Loads GTA bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
width = 1920
height = 1080
info = self.dataset[self.image_id_at(index)]
labels = info['labels'] # a list of dict
# get the kitti part out and insert the tracking id
boxes = ds.get_box2d_array(labels).astype(float)[:, :4]
tid = ds.get_label_array(labels, ['id'], (0)).astype(int)
num_objs = len(tid)
#gt_cls = ds.get_label_array(labels, ['class'], (0))
gt_cls = np.array(['foreground'] * num_objs)
gt_classes = np.ones(num_objs)
# actually just one single value,
ignore = ds.get_label_array(labels, ['attributes', 'ignore'],
(0)).astype(int)
cam_calib = np.array(info['intrinsics']['cali'])
location = ds.get_label_array(labels, ['box3d', 'location'],
(0, 3)).astype(float)
ext_loc = np.hstack([location, np.ones([len(location), 1])]) # (B, 4)
proj_loc = ext_loc.dot(cam_calib.T) # (B, 4) dot (3, 4).T => (B, 3)
center = proj_loc[:, :2] / proj_loc[:, 2:3] # normalize
seg_areas = (boxes[:, 2] - boxes[:, 0] + 1) * \
(boxes[:, 3] - boxes[:, 1] + 1)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
endvid = np.zeros((num_objs), dtype=np.uint16)
# pad to make it consistent
if self.endvid[self.image_id_at(index)]:
endvid += 1
for ix in range(num_objs):
cls = self._class_to_ind[gt_cls[ix].strip()]
gt_classes[ix] = cls
overlaps[ix, cls] = 1.0
ds.validate_boxes(boxes, width=width, height=height)
print("overlaps:", overlaps)
overlaps = scipy.sparse.csr_matrix(overlaps)
# print("end of video:", endvid)
# print("overlaps:", overlaps)
info_set = {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False,
'seg_areas': seg_areas,
'ignore': ignore,
'end_vid': endvid,
'center': center
}
# print(info_set)
return info_set
def _load_coco_annotation(self, index):
"""
Loads COCO bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
# print('index:\n', index)
im_ann = self._COCO.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
# catIds = self._COCO.getCatIds(catNms=['cell phone', 'laptop'])
# Change here to load only boxes that correspond to cell & laptot ? or load everything and call them "other"
annIds = self._COCO.getAnnIds(imgIds=index,
catIds=self._c_to_keep,
iscrowd=None)
# print('annIds size:\n', len(annIds))
objs = self._COCO.loadAnns(annIds)
# print('objs:\n', objs)
# Sanitize bboxes -- some are invalid
valid_objs = []
for obj in objs:
x1 = np.max((0, obj['bbox'][0]))
y1 = np.max((0, obj['bbox'][1]))
x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
valid_objs.append(obj)
objs = valid_objs
self._total_positive += 1 if len(
objs) >= 1 else 0 # <-------------- ADDED
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
# Lookup table to map from COCO category ids to our internal class
# indices
############################################################################### lookup table
# <------------------------------------------------------------------------------------------ Modified
coco_cat_id_to_class_ind = dict([(self._class_to_coco_cat_id[cls],
self._class_to_ind[cls])
for cls in self._classes[1:]])
#coco_cat_id_to_class_ind = dict([(c, i+1) for i, c in enumerate(self._c_to_keep)])
# coco_cat_id_to_class_ind = {73:1, 77:2}
#print('coco_cat_id_to_class_ind:\n', coco_cat_id_to_class_ind)
for ix, obj in enumerate(objs):
cls = coco_cat_id_to_class_ind[obj['category_id']]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
seg_areas[ix] = obj['area']
if obj['iscrowd']:
# Set overlap to -1 for all classes for crowd objects
# so they will be excluded during training
overlaps[ix, :] = -1.0
else:
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False,
'seg_areas': seg_areas
}
def _load_openimages_annotation(self, i):
"""
Loads openimages bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
width = self._imsize[i][1]
height = self._imsize[i][2]
index = self._image_index[i]
if index not in self._annotation_dict:
objs = []
print("not loading image {0}".format(index))
else:
objs = self._annotation_dict[index]
# Sanitize bboxes -- some are invalid
valid_objs = []
for number, obj in enumerate(objs):
'''
x1 = np.max((0, int(width * obj[2])))
x2 = np.min((width - 1, int((width - 1) * obj[3])))
y1 = np.max((0, int(height * obj[4])))
y2 = np.min((height - 1, int((height - 1) * obj[5])))
'''
x1 = int(width * obj[2])
x2 = int((width - 1) * obj[3])
y1 = int(height * obj[4])
y2 = int((height - 1) * obj[5])
# TODO are the BB cooridnates diferent?
if x2 >= x1 and y2 >= y1:
valid_objs.append((obj[0], obj[1], x1, y1, x2, y2))
else:
print("filtered")
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
# Lookup table to map from COCO category ids to our internal class
# indices
coco_cat_id_to_class_ind = dict([(self._class_to_coco_cat_id[cls],
self._class_to_ind[cls])
for cls in self._classes[1:]])
for ix, obj in enumerate(objs):
cls = coco_cat_id_to_class_ind[obj[1]]
boxes[ix, :] = [obj[2], obj[3], obj[4], obj[5]]
gt_classes[ix] = cls
seg_areas[ix] = (obj[4] - obj[2]) * (obj[5] - obj[3])
# TODO what does overlaps do?
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False,
'seg_areas': seg_areas
}
def _load_coco_annotation(self, index):
"""
Loads COCO bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
# print type(index)
im_ann = self._COCO.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
img_name = im_ann['file_name']
annIds = self._COCO.getAnnIds(imgIds=index, iscrowd=None)
objs = self._COCO.loadAnns(annIds)
# Sanitize bboxes -- some are invalid
valid_objs = []
for obj in objs:
x1 = np.max((0, obj['bbox'][0]))
y1 = np.max((0, obj['bbox'][1]))
x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
valid_objs.append(obj)
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
for ix, obj in enumerate(objs):
cls = self._coco_ind_to_class_ind[obj['category_id']]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
seg_areas[ix] = obj['area']
if obj['iscrowd']:
# Set overlap to -1 for all classes for crowd objects
# so they will be excluded during training
overlaps[ix, :] = -1.0
else:
overlaps[ix, cls] = 1.0
# check bbox boundary validation
ds_utils.validate_boxes(boxes, width=width, height=height)
# max overlap with gt over classes (columns)
max_overlaps = overlaps.argmax(axis=1)
# gt class that had the max overlap
max_classes = overlaps.argmax(axis=1)
# sanity checks
# max overlap of 0 => class should be zero (background)
zero_inds = np.where(max_overlaps == 0)[0]
assert all(max_classes[zero_inds] == 0)
# max overlap > 0 => class should not be zero (must be a fg class)
nonzero_inds = np.where(max_overlaps > 0)[0]
assert all(max_classes[nonzero_inds] != 0)
return {'image': img_name,
'height': height,
'width': width,
'boxes' : boxes,
'gt_classes': gt_classes,
'gt_overlaps' : overlaps,
'max_classes': max_classes,
'max_overlaps': max_overlaps,
'flipped' : False,
'seg_areas' : seg_areas}
def _loadCarlaAnnotation(self, index, vis=False):
"""
Load the annotations of the ith image from the info list
"""
width, height = 1920, 1080
# "labels"
info = self.infos[self.image_id_at(index)]
labels = info["labels"]
# Object ids
objIds = ds.get_label_array(labels, key_list=["id"],
empty_shape=(0, )).astype(int)
nObjects = len(objIds)
# Ignores
ignores = ds.get_label_array(labels,
key_list=["attributes", "ignore"],
empty_shape=(0, )).astype(int)
# Ground-truth class
objClasses = ds.getClassArray(
labels, self.classes[0], emptyShape=(0, )) # Class name of objects
gtClassses = np.array([self.cls2Ind[c] for c in objClasses
]) # convert to the index of class
# Boxes
boxes = ds.get_box2d_array(labels)[:, :4] # ignore the confidence
# 3d object centers
locations = ds.get_label_array(labels,
key_list=["box3d", "location"],
empty_shape=(0, 3)).astype(float)
centers = self._projection(locations)
# Overlaps (one-hot vector form of labels)
overlaps = np.zeros((nObjects, self.num_classes), dtype=float)
overlaps[np.arange(nObjects), gtClassses] += 1
overlaps = scipy.sparse.csr_matrix(
overlaps) # convert to a sparse matrix
# Area of segments
segAreas = self._computeSegAreas(boxes)
# End-of-video id (flag)
endvids = np.zeros(nObjects, dtype=int)
if (self.endvids[self.image_id_at(index)]):
# Last image of the video, set them to 1
endvids += 1
# Validation boxes
flag = ds.validate_boxes(boxes, width=width, height=height)
# Information set
infoSet = None
if (flag):
# Debug
# print(self.image_path_at(index))
# print(endvids)
if vis:
im = cv2.imread(self.image_path_at(index))
im2show = np.copy(im)
im2show = self._drawBBoxAndCenter(im2show, boxes, centers)
cv2.imwrite(
os.path.join("vis", "inputs",
'result%d.png' % (index + 1)), im2show)
infoSet = {
"width": width,
"height": height,
"boxes": boxes,
"gt_classes": gtClassses,
"gt_overlaps": overlaps,
"flipped": False,
"seg_areas": segAreas,
"ignore": ignores,
"end_vid": endvids,
"center": centers
}
return infoSet
def _load_coco_annotation(self, index):
"""
Loads COCO bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
im_ann = self._COCO.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
annIds = self._COCO.getAnnIds(imgIds=index, iscrowd=None)
objs = self._COCO.loadAnns(annIds)
'''
annotation{
"id": int,
"image_id": int,
"category_id": int,
"segmentation": RLE or [polygon],
"area": float,
"bbox": [x,y,width,height],
"iscrowd": 0 or 1,
}
'''
############################ write information for debugging
# f = open("./tools/tracking_file.txt", "w")
# f.write("\nimage index: " + str(index) + "\n")
# f.write("\nimage height, width: " + str(height) + " " + str(width) + "\n")
############################
# Sanitize bboxes -- some are invalid
valid_objs = []
count = 0
for obj in objs:
x1 = np.max((0, obj['bbox'][0])) #col
y1 = np.max((0, obj['bbox'][1])) #row
x2 = np.min((width - 1, x1 + np.max(
(0, obj['bbox'][2] - 1)))) #col
y2 = np.min((height - 1, y1 + np.max(
(0, obj['bbox'][3] - 1)))) #row
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
################# get segmentation mask of this object###############################################
if cfg.TRAIN.MASK_REG:
count = count + 1
obj['seg_mask_ind'] = [
index, count
] #image index and order of this object in the image
seg_mask = self._COCO.annToMask(obj)
seg_mask_save = np.zeros((height, width), dtype=np.bool_)
seg_mask_save = seg_mask
seg_mask_save = seg_mask_save.astype(bool)
# save the mask into disk
#seg_mask_path = './data/cache/seg_mask_coco_gt/' + str(index) + '_' + str(count) + '_segmask.sm'
seg_mask_path = './data/cache/' + 'GTsegmask_' + cfg.TRAIN.TRAINING_DATA + '/' + str(
index) + '_' + str(count) + '_segmask.sm'
with open(seg_mask_path, 'wb') as f_seg_save:
cPickle.dump(seg_mask_save, f_seg_save,
cPickle.HIGHEST_PROTOCOL)
# f.write("\nobject: " + str(count) + "\n")
# f.write("\ncoordinate (x1(col) y1(row) x2 y2): " + str(x1) + " " + str(y1) + " " + str(x2) + " " + str(y2) + "\n")
# f.write("\narea: " + str(obj['area']) + "\n")
# f.write("\nsegmentation mask size: " + str(seg_mask.shape[0]) + " " + str(seg_mask.shape[1]) + "\n")
# f.write("\n")
# print segmentation mask for debugging
# if index == 262145 and count == 1:
# f_seg = open("./tools/segment_mask.txt", "w")
# for i1 in xrange(0, height):
# for i2 in xrange(0, width):
# f_seg.write(str(seg_mask[i1][i2]) + " ")
# f_seg.write("\n")
# f_seg.close()
######################################################################################################
valid_objs.append(obj)
# f.write("\n")
# f.close()
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
################
seg_mask_inds = np.zeros(
(num_objs, 2),
dtype=np.uint32) #save image index and object index in this image
#seg_masks = np.zeros((num_objs, height * width), dtype=np.bool_) #bool data type
################
# Lookup table to map from COCO category ids to our internal class
# indices
coco_cat_id_to_class_ind = dict([(self._class_to_coco_cat_id[cls],
self._class_to_ind[cls])
for cls in self._classes[1:]])
for ix, obj in enumerate(
objs
): #moi element cua objs la 1 obj voi nhieu fields nhu 'category_id','clean_bbox',.... Chu ban than objs ko co cac fields nhu tren. objs chi chua obj
cls = coco_cat_id_to_class_ind[obj['category_id']]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
seg_areas[ix] = obj['area']
###########
if cfg.TRAIN.MASK_REG:
#seg_masks[ix] = obj['seg_mask']
seg_mask_inds[ix, :] = obj['seg_mask_ind']
###########
if obj['iscrowd']:
# Set overlap to -1 for all classes for crowd objects
# so they will be excluded during training
overlaps[ix, :] = -1.0
else:
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
############
#seg_masks = scipy.sparse.csr_matrix(seg_masks) #store seg_masks as sparse matrix
############
if not cfg.TRAIN.MASK_REG:
return {
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False, #--->flag day la anh original
'seg_areas': seg_areas
}
else:
#print "========================Have seg_mask_inds======================"
return {
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False, #--->flag day la anh original
'seg_areas': seg_areas,
'seg_mask_inds': seg_mask_inds
}
def _load_coco_annotation(self, index):
"""
Loads COCO bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
im_ann = self._COCO.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
annIds = self._COCO.getAnnIds(imgIds=index, iscrowd=None)
objs = self._COCO.loadAnns(annIds)
# Sanitize bboxes -- some are invalid
valid_objs = []
for obj in objs:
x1 = np.max((0, obj['bbox'][0]))
y1 = np.max((0, obj['bbox'][1]))
x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
# mask = self._COCO.annToMask(obj)
# start_h = np.round(np.max((1, y1))).astype(np.int)
# end_h = np.round(np.min((height, y2))).astype(np.int)
# start_w = np.round(np.max((1, x1))).astype(np.int)
# end_w = np.round(np.min((width, x2))).astype(np.int)
# # print 'mask shape---',mask.shape
# # print '\nsize:',start_h,end_h,start_w,end_w
# # print '\n,x1,y1,x2,y2',x1,y1,x2,y2
# if start_h==end_h:
# end_h +=1
# if start_w == end_w:
# end_w +=1
# cropped_mask = mask[start_h:end_h, start_w:end_w]
# resized_mask = cv2.resize(cropped_mask, (cfg.MASK_SIZE, cfg.MASK_SIZE), interpolation=cv2.INTER_CUBIC)
obj['mask'] = obj
# valid_objs.append(obj)
if obj['iscrowd'] == 0:
valid_objs.append(obj)
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
masks = []
# masks = np.zeros((num_objs,cfg.MASK_SIZE,cfg.MASK_SIZE),dtype=np.float32)
# Lookup table to map from COCO category ids to our internal class
# indices
coco_cat_id_to_class_ind = dict([(self._class_to_coco_cat_id[cls],
self._class_to_ind[cls])
for cls in self._classes[1:]])
for ix, obj in enumerate(objs):
cls = coco_cat_id_to_class_ind[obj['category_id']]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
seg_areas[ix] = obj['area']
# masks[ix,:,:] = obj['mask']
masks.append(obj['mask'])
if obj['iscrowd']:
# Set overlap to -1 for all classes for crowd objects
# so they will be excluded during training
overlaps[ix, :] = -1.0
else:
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False,
'seg_areas': seg_areas,
'masks': masks
}
def _load_kitti_annotation(self, index):
"""
Loads KITTI bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
im_ann = self._KITTI.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
# Follow 'demo_load_kitti_dataset.py by Soonmin'
hRng, occLevel, tRng = self.config['hRng'], self.config[
'occLevel'], self.config['truncRng']
# Load annotation ids
annIds = self._KITTI.getAnnIds(imgIds=index,
catIds=self._raw_cat_ids,
hRng=hRng,
occLevel=occLevel,
truncRng=tRng)
#annIds = self._KITTI.getAnnIds(imgIds=index, hRng=hRng, occLevel=occLevel, truncRng=tRng)
objs = self._KITTI.loadAnns(annIds)
# Sanitize bboxes -- some are invalid
valid_objs = []
for obj in objs:
x1 = np.max((0, obj['bbox'][0]))
y1 = np.max((0, obj['bbox'][1]))
x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
# All valid annotations must satisfy below condition
if x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
valid_objs.append(obj)
objs = valid_objs
num_objs = len(objs)
# In traffic scene datasets (e.g. KITTI, KAIST),
# some images may not contain any target object instance.
# Then, num_objs == 0.
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
# Lookup table to map from KITTI category ids to our internal class indices
kitti_cat_id_to_class_ind = dict([(self._class_to_kitti_cat_id[cls],
self._class_to_ind[cls])
for cls in self._classes[1:]])
for ix, obj in enumerate(objs):
cls = kitti_cat_id_to_class_ind[obj['category_id']]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False, # Data augmentation
'gamma': False, # Data augmentation
'crop': None, # Data augmentation
'jitter': False
} # Data augmentation
def _load_coco_annotation(self, index):
"""
Loads COCO bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
im_ann = self._COCO.loadImgs(index)[0]
im_path = self.image_path_from_index(index)
width = im_ann['width']
height = im_ann['height']
# Get the useful information
# reference = self.reference_image[index]
# save_seq = reference.keys()
annIds = self._COCO.getAnnIds(imgIds=index, iscrowd=None)
objs = self._COCO.loadAnns(annIds)
# Sanitize bboxes -- some are invalid
valid_objs = []
for i, obj in enumerate(objs):
x1 = np.max((0, obj['bbox'][0]))
y1 = np.max((0, obj['bbox'][1]))
x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
valid_objs.append(obj)
# if i in save_seq:
entry = {'boxes': obj['clean_bbox'], 'image_path': im_path}
self.cat_data[self.coco_cat_id_to_class_ind[
obj['category_id']]].append(entry)
objs = valid_objs
num_objs = len(objs)
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
for ix, obj in enumerate(objs):
cls = self.coco_cat_id_to_class_ind[obj['category_id']]
boxes[ix, :] = obj['clean_bbox']
gt_classes[ix] = cls
seg_areas[ix] = obj['area']
if obj['iscrowd']:
# Set overlap to -1 for all classes for crowd objects
# so they will be excluded during training
overlaps[ix, :] = -1.0
else:
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False,
'seg_areas': seg_areas
}
def _load_hico_annotation(self, index):
"""
Loads COCO bounding-box instance annotations. Crowd instances are
handled by marking their overlaps (with all categories) to -1. This
overlap value means that crowd "instances" are excluded from training.
"""
#im_ann = self._COCO.loadImgs(index)[0]#----!!!
#print(f'label:{index}' )
im_ann = self._HICO[index] #----!!!! (0~N-1)#---!?!?!?
width = im_ann['width']
height = im_ann['height']
#annIds = self._COCO.getAnnIds(imgIds=index, iscrowd=None)#----!!!
#objs = self._COCO.loadAnns(annIds)#----!!!
objs = im_ann['ann'] #----!!!
# Sanitize bboxes -- some are invalid
valid_objs = []
clean_bbox = []
clean_label = []
for ix in range(len(objs['bboxes'])): #----!!!
#for obj in objs:
x1 = np.max((0, objs['bboxes'][ix][0]))
y1 = np.max((0, objs['bboxes'][ix][1]))
x2 = np.min((width - 1, x1 + np.max(
(0, objs['bboxes'][ix][2] - 1))))
y2 = np.min((height - 1, y1 + np.max(
(0, objs['bboxes'][ix][3] - 1))))
if x2 >= x1 and y2 >= y1 and x2 < width - 1 and y2 < height - 1:
clean_bbox.append([x1, y1, x2, y2])
clean_label.append(objs['labels'][ix])
objs['clean_bbox'] = clean_bbox
objs['clean_label'] = clean_label
num_objs = len(objs['clean_bbox']) #len(objs)#---!!!
boxes = np.zeros((num_objs, 4), dtype=np.uint16)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
seg_areas = np.zeros((num_objs), dtype=np.float32)
# Lookup table to map from COCO category ids to our internal class
# indices
#coco_cat_id_to_class_ind = dict([(self._class_to_coco_cat_id[cls],
# self._class_to_ind[cls])
# for cls in self._classes[1:]])
for ix in range(num_objs): #----!!!
#for ix, obj in enumerate(objs):
#cls = coco_cat_id_to_class_ind[obj['category_id']]#----!!!
cls = objs['clean_label'][ix] #---!!!
boxes[ix, :] = objs['clean_bbox'][ix] #obj['clean_bbox']#----!!!
gt_classes[ix] = cls
#seg_areas[ix] = obj['area']
#if obj['iscrowd']:
# # Set overlap to -1 for all classes for crowd objects
# # so they will be excluded during training
# overlaps[ix, :] = -1.0
#else:
overlaps[ix, cls] = 1.0
ds_utils.validate_boxes(boxes, width=width, height=height)
overlaps = scipy.sparse.csr_matrix(overlaps)
return {
'width': width,
'height': height,
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'flipped': False,
'seg_areas': seg_areas
}
