def main():
logdir = os.path.dirname(args.seg_dir)
log_filename = os.path.join(logdir, 'corloc_eval_%s.log' % datetime.now())
logging.basicConfig(filename=log_filename, level=logging.DEBUG)
tp = np.zeros(len(pascal_classes))
gt = np.zeros(len(pascal_classes))
image_list = get_list_from_file(args.image_list)
for image in image_list:
_, img_labels, bboxes = get_annotations_from_xml(
os.path.join(args.annotation_dir, '%s.xml' % image),
pascal_classes)
gt += np.maximum(img_labels, 0)
seg_mask = cv2.imread(os.path.join(args.seg_dir, '%s.png' % image))
seg_mask = seg_mask[:, :, 0]
seg_mask = seg_mask.astype(np.int32)
for i in range(1, len(pascal_classes)):
if img_labels[i] > 0:
mask = (seg_mask == i)
if np.sum(mask) > 0:
regionmap = label(mask)
region_prop = regionprops(regionmap)
region_area = np.array([r.area for r in region_prop])
best_region_idx = np.argmax(region_area)
bbox = np.array(region_prop[best_region_idx].bbox)
bbox = np.expand_dims(bbox, 0)
gt_bbox = bboxes[bboxes[:, 0] == i][:, 1:]
iou = get_iou_matrix(gt_bbox, bbox)
if np.amax(iou > 0.5):
tp[i] += 1
for i in range(1, len(pascal_classes)):
logging.info('%s: %.3f' % (pascal_classes[i], tp[i] / gt[i]))
logging.info('CorLoc: %.3f' % (np.sum(tp[1:] / gt[1:]) /
(len(pascal_classes) - 1)))
logging.info('Instance correct rate: %d/%d' % (np.sum(tp), np.sum(gt)))
def main():
logdir = os.path.dirname(args.heatmap_dir)
log_filename = os.path.join(logdir, 'pointing_eval_%s.log' % datetime.now())
logging.basicConfig(filename=log_filename, level=logging.DEBUG)
logging.info('Folder to be evaluated: %s'%args.heatmap_dir)
tp = np.zeros(len(pascal_classes))
gt = np.zeros(len(pascal_classes))
image_list = get_list_from_file(args.image_list)
for image in image_list:
img_size, img_labels, bboxes = get_annotations_from_xml(os.path.join(args.annotation_dir, '%s.xml'%image),
pascal_classes)
gt += np.maximum(img_labels, 0)
for i in range(1, len(pascal_classes)):
if img_labels[i] > 0:
heatmap = np.load(os.path.join(args.heatmap_dir, '%s_%s.npy'%(image, pascal_classes[i])))
heatmap = cv2.resize(heatmap, (img_size[1], img_size[0]))
global_argmax = np.where(heatmap==np.amax(heatmap))
hmax = global_argmax[0]
wmax = global_argmax[1]
point = np.concatenate([hmax, wmax]) + 1
gt_bbox = bboxes[bboxes[:, 0] == i][:, 1:]
is_in = point_in_box(point, gt_bbox)
if np.amax(is_in):
tp[i] += 1
for i in range(1, len(pascal_classes)):
logging.info('%s: %.3f'%(pascal_classes[i], tp[i]/gt[i]))
logging.info('CorLoc: %.3f'%(np.sum(tp[1:]/gt[1:])/(len(pascal_classes) - 1)))
logging.info('Instance correct rate: %d/%d'%(np.sum(tp), np.sum(gt)))
def test_get_label_and_objects_from_xml():
xmlpath = '%s/data/utils/test_xml.xml' % ROOT_DIR
pascal_classes = [
'__background__', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle',
'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse',
'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train',
'tvmonitor'
]
size, labels, objects = file_utils.get_annotations_from_xml(
xmlpath, pascal_classes)
assert np.array_equal(size, np.array([375, 500, 3]))
assert np.array_equal(
labels,
np.array(
[0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0]))
assert np.array_equal(
objects,
np.array([[9, 211, 263, 339, 324], [7, 264, 165, 372, 253],
[-9, 244, 5, 374, 67], [9, 194, 241, 299, 295],
[-15, 186, 277, 220, 312]]))
def main(unused):
net_CAM = caffe.Net(
os.path.join(cfg.root_dir, 'SEC', 'deploy.prototxt'),
os.path.join(cfg.root_dir, 'SEC', 'weights.caffemodel'), caffe.TEST)
data_dir = getattr(data.cfg, '%s_dir' % args.dataset)
filelist = get_list_from_file(
os.path.join(data_dir, 'ImageSets', 'Main', '%s.txt' % args.image_set))
print 'Num of images', len(filelist)
for file in filelist:
image = pylab.imread(
os.path.join(data_dir, 'JPEGImages', '%s.jpg' % file))
H = image.shape[0]
W = image.shape[1]
_, labels, _ = get_annotations_from_xml(
os.path.join(data_dir, 'Annotations', '%s.xml' % file),
data.cfg.pascal_classes)
if np.sum(np.absolute(labels[1:])) > 0:
print file
net_CAM.blobs['images'].data[...][0] = preprocess(image, 321.0)
net_CAM.forward()
CAM_scores = net_CAM.blobs['fc7_CAM'].data[0]
params = net_CAM.params['scores'][0].data[...]
for i in range(1, 21):
if labels[i] != 0:
w = params[i - 1]
heat_maps = np.sum(CAM_scores * w[:, None, None], axis=0)
if args.result_dir:
np.save(
'%s/%s_%s' % (args.result_dir, file,
data.cfg.pascal_classes[i]),
heat_maps)
if args.vis_dir:
vis = visualize_heatmap(np.maximum(heat_maps, 0))
cv2.imwrite(
'%s/%s_%s.png' %
(args.vis_dir, file, data.cfg.pascal_classes[i]),
vis)
def main(unused):
category_list = file_utils.get_list_from_file(FLAGS.category_list_file)
num_categories = len(category_list)
co_occur_freq = np.zeros((num_categories, num_categories))
category_freq = np.zeros((num_categories))
image_list = file_utils.get_list_from_file(FLAGS.image_list)
for image in image_list:
xml_path = os.path.join(FLAGS.annotation_dir, '%s.xml' % image)
_, labels, objs = file_utils.get_annotations_from_xml(
xml_path, category_list)
obj_labels = np.unique(np.absolute(objs[:, 0]))
labels = np.absolute(labels)
if obj_labels.shape[0] == 1:
co_occur_freq[obj_labels[0], obj_labels[0]] += 1
for l in range(0, obj_labels.shape[0]):
for k in range(l + 1, obj_labels.shape[0]):
co_occur_freq[obj_labels[l], obj_labels[k]] += 1
co_occur_freq[obj_labels[k], obj_labels[l]] += 1
category_freq += labels
co_occur_prob = co_occur_freq / np.expand_dims(category_freq, 1)
if FLAGS.save_dir:
np.save(os.path.join(FLAGS.save_dir, 'pascal_co_occur'), co_occur_prob)
def main():
logdir = os.path.dirname(args.seg_dir)
log_filename = os.path.join(logdir, 'seed_eval_%s.log' % datetime.now())
logging.basicConfig(filename=log_filename, level=logging.DEBUG)
logging.info('Folder to be evaluated: %s, %s'%(args.seg_dir, args.seedmask_dir))
tp = np.zeros(len(pascal_classes))
gt = np.zeros(len(pascal_classes))
pd = np.zeros(len(pascal_classes))
image_list = get_list_from_file(args.image_list)
for image in image_list:
img_size, img_labels, bboxes = get_annotations_from_xml(os.path.join(args.annotation_dir, '%s.xml'%image),
pascal_classes)
fg_map = np.zeros(img_size, dtype=np.bool)
for i in range(1, len(pascal_classes)):
if img_labels[i] > 0:
seg = cv2.imread(os.path.join(args.seg_dir, '%s_%s.png'%(image, pascal_classes[i])))
seedmask = cv2.imread(os.path.join(args.seedmask_dir, '%s_%s.png'%(image, pascal_classes[i])))
cue = np.logical_and((seg == i), (seedmask==1))
gt_bbox = bboxes[np.absolute(bboxes[:, 0]) == i][:, 1:]
box_map = boxes_to_seg(gt_bbox, img_size)
gt[i] += np.sum(box_map)
tp[i] += np.sum(np.logical_and(box_map, cue))
pd[i] += np.sum(cue)
fg_map = np.logical_or(fg_map, box_map)
bg_cue = np.logical_and((seg == 0), (seedmask==1))
bg_map = np.logical_not(fg_map)
gt[0] += np.sum(bg_map)
tp[0] += np.sum(np.logical_and(bg_map, bg_cue))
pd[0] += np.sum(bg_cue)
for i in range(len(pascal_classes)):
logging.info('%s: recall: %.3f, precision: %.3f'%(pascal_classes[i], tp[i]/gt[i], tp[i]/pd[i]))
logging.info('Average recall: %.3f'%(np.sum(tp[1:]/gt[1:])/(len(pascal_classes) - 1)))
logging.info('Average precision: %.3f'%(np.sum(tp[1:]/pd[1:])/(len(pascal_classes) - 1)))
def _find_image_files(data_dir, annotation_dir, image_list_file, category_list_file):
"""Build a list of all images files and labels in the data set.
Args:
data_dir: string, path to the root directory of images.
annotation_dir: path to annotation files
image_list_file: a text file listing all image file names
category_list_file: a text file listing all possible classes for this dataset
Returns:
filenames: list of strings; each string is a path to an image file.
labels: list of vector; each vector identifies the ground truth.
roi_labels: list of vector, identifying one class of interest
obj_labels: list of array, instance-level annotation
obj_bboxes: list of array, instance-level box annotation
proposals: regions of interest (for the class of interest if roi_labels are valid,
else for all exsiting categories
"""
labels = []
filenames = []
proposals = []
roi_labels = []
obj_labels = []
obj_bboxes = []
filename_list = file_utils.get_list_from_file(image_list_file)
class_list = file_utils.get_list_from_file(category_list_file)
print(len(filename_list))
# Construct the list of JPEG files and labels.
for filename in filename_list:
jpeg_file_path = '%s/%s.jpg' % (data_dir, filename)
# read .xml label file
xml_path = '%s/%s.xml' % (annotation_dir, filename)
img_size, label, objects = file_utils.get_annotations_from_xml(xml_path, class_list)
bbox = objects[:, 1:]
obj_label = objects[:, 0]
if FLAGS.normalize_bbox:
bbox = box_utils.bbox_normalize(bbox, img_size[0:2], one_index=True)
# get bbox
if FLAGS.roi_dir:
roi_dir = os.path.join(FLAGS.roi_dir, '%s.mat' % filename)
roi_contents = sio.loadmat(roi_dir)
if not 'boxes' in roi_contents: # category-dependent proposals
file_duplicates = 0
for key in roi_contents.keys():
if not key.startswith('__'):
label_index = class_list.index(key)
if label[label_index] == -1:
continue
roi_label = np.zeros(len(class_list), dtype=np.int64)
roi_label[label_index] = 1
proposal = roi_contents[key]
assert proposal.shape[1] == 4
if FLAGS.normalize_bbox:
proposal = box_utils.bbox_normalize(proposal, img_size[0:2], one_index=True)
file_duplicates += 1
proposals.append(proposal)
roi_labels.append(roi_label)
else: # category-independent proposals
file_duplicates = 1
proposal = roi_contents['boxes']
assert proposal.shape[1] == 4
if FLAGS.normalize_bbox:
proposal = box_utils.bbox_normalize(proposal, img_size[0:2], one_index=True)
proposals.append(proposal)
roi_labels.append(label)
else:
file_duplicates = 1
proposals.append(np.array([[-1, -1, -1, -1]]))
roi_labels.append(label)
labels.extend([label] * file_duplicates)
filenames.extend([jpeg_file_path]*file_duplicates)
obj_labels.extend([obj_label]*file_duplicates)
obj_bboxes.extend([bbox]*file_duplicates)
print(len(labels))
print(len(filenames))
print(len(obj_labels))
shuffled_index = range(len(filenames))
random.seed(12345)
random.shuffle(shuffled_index)
filenames = [filenames[i] for i in shuffled_index]
labels = [labels[i] for i in shuffled_index]
roi_labels = [roi_labels[i] for i in shuffled_index]
proposals = [proposals[i] for i in shuffled_index]
obj_labels = [obj_labels[i] for i in shuffled_index]
obj_bboxes = [obj_bboxes[i] for i in shuffled_index]
# print(filenames[0])
# print(labels[0])
# print(roi_labels[0])
# print(obj_bboxes[0])
# print(obj_labels[0])
return filenames, labels, roi_labels, proposals, obj_labels, obj_bboxes
def _find_image_files(data_dir, annotation_dir, image_list_file,
category_list_file, seg_dir, seg_mask_dir):
"""Build a list of all images files and labels in the data set."""
labels = []
filenames = []
seg_files = []
roi_labels = []
seg_mask_files = []
filename_list = file_utils.get_list_from_file(image_list_file)
class_list = file_utils.get_list_from_file(category_list_file)
print(len(filename_list))
# Construct the list of JPEG files and labels.
for filename in filename_list:
jpeg_file_path = '%s/%s.jpg' % (data_dir, filename)
# read .xml label file
xml_path = '%s/%s.xml' % (annotation_dir, filename)
img_size, label, _ = file_utils.get_annotations_from_xml(
xml_path, class_list)
# get segmentation
if FLAGS.one_hot:
file_duplicates = 0
for i in range(1, len(class_list)):
cls = class_list[i]
if label[i] > 0:
file_duplicates += 1
roi_label = np.zeros(len(class_list), dtype=np.int64)
roi_label[i] = 1
seg_file = os.path.join(seg_dir,
'%s_%s.png' % (filename, cls))
seg_mask_file = os.path.join(seg_mask_dir,
'%s_%s.png' % (filename, cls))
seg_files.append(seg_file)
seg_mask_files.append(seg_mask_file)
roi_labels.append(roi_label)
else:
file_duplicates = 1
roi_label = label
seg_file = os.path.join(seg_dir, '%s.png' % (filename))
seg_mask_file = os.path.join(seg_mask_dir, '%s.png' % (filename))
seg_files.append(seg_file)
seg_mask_files.append(seg_mask_file)
roi_labels.append(roi_label)
labels.extend([label] * file_duplicates)
filenames.extend([jpeg_file_path] * file_duplicates)
# Shuffle the ordering of all image files in order to guarantee
# random ordering of the images with respect to label in the
# saved TFRecord files. Make the randomization repeatable.
print(len(labels))
print(len(filenames))
print(len(seg_files))
print(len(seg_mask_files))
shuffled_index = range(len(filenames))
random.seed(12345)
random.shuffle(shuffled_index)
filenames = [filenames[i] for i in shuffled_index]
labels = [labels[i] for i in shuffled_index]
roi_labels = [roi_labels[i] for i in shuffled_index]
seg_files = [seg_files[i] for i in shuffled_index]
seg_mask_files = [seg_mask_files[i] for i in shuffled_index]
# print(filenames[0])
# print(labels[0])
# print(roi_labels[0])
# print(obj_bboxes[0])
# print(obj_labels[0])
# return
return filenames, labels, roi_labels, seg_files, seg_mask_files