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_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 = []
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)]
# 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, :]
box_list.append(boxes)
# Sanity check
im_ann = self._COCO.loadImgs(index)[0]
width = im_ann['width']
height = im_ann['height']
ds_utils.validate_boxes(boxes, width=width, height=height)
return self.create_roidb_from_box_list(box_list, gt_roidb)