def get_neg_windows(self, num, cls=None, window_params=None, max_overlap=0, max_num_images=250):
"""
Return array of num windows that can be generated with window_params
that do not overlap with ground truth by more than max_overlap.
* If cls is not given, returns ground truth for all classes.
* If max_num_images is given, samples from at most that many images.
"""
sw = SlidingWindows(self, self)
if not window_params:
window_params = sw.get_default_window_params(cls)
all_windows = []
image_inds = self.get_pos_samples_for_class(cls)
max_num = len(image_inds)
inds = image_inds
if max_num_images:
inds = ut.random_subset(image_inds, max_num_images)
num_per_image = round(1.0 * num / max_num)
for ind in inds:
image = self.images[ind]
windows = image.get_windows(window_params)
gts = image.get_ground_truth(cls)
for gt in gts.arr:
overlaps = BoundingBox.get_overlap(windows[:, :4], gt[:4])
windows = windows[overlaps <= max_overlap, :]
if windows.shape[0] == 0:
continue
ind_to_take = ut.random_subset_up_to_N(windows.shape[0], num_per_image)
all_windows.append(np.hstack((windows[ind_to_take, :], np.tile(ind, (ind_to_take.shape[0], 1)))))
all_windows = np.concatenate(all_windows, 0)
return all_windows[:num, :]
def get_pos_windows(self, cls=None, window_params=None, min_overlap=0.7):
"""
Return array of all ground truth windows for the class, plus windows
that can be generated with window_params that overlap with it by more
than min_overlap.
* If cls not given, return positive windows for all classes.
* If window_params not given, use default for the class.
* Adjust min_overlap to fetch fewer windows.
"""
sw = SlidingWindows(self, self)
if not window_params:
window_params = sw.get_default_window_params(cls)
overlapping_windows = []
image_inds = self.get_pos_samples_for_class(cls)
times = []
window_nums = []
for i in image_inds:
image = self.images[i]
gts = image.get_ground_truth(cls)
if gts.arr.shape[0] > 0:
overlap_wins = gts.arr[:, :4]
overlap_wins = np.hstack(
(overlap_wins, np.tile(i, (overlap_wins.shape[0], 1))))
overlapping_windows.append(overlap_wins.astype(int))
windows, time_elapsed = image.get_windows(window_params,
with_time=True)
window_nums.append(windows.shape[0])
times.append(time_elapsed)
for gt in gts.arr:
overlaps = BoundingBox.get_overlap(windows[:, :4], gt[:4])
overlap_wins = windows[overlaps >= min_overlap, :]
overlap_wins = np.hstack(
(overlap_wins, np.tile(i, (overlap_wins.shape[0], 1))))
overlapping_windows.append(overlap_wins.astype(int))
windows = windows[overlaps < min_overlap, :]
overlapping_windows = np.concatenate(overlapping_windows, 0)
print(
"Windows generated per image: %d +/- %.3f, in %.3f +/- %.3f sec" %
(np.mean(window_nums), np.std(window_nums), np.mean(times),
np.std(times)))
return overlapping_windows
def get_neg_windows(self,
num,
cls=None,
window_params=None,
max_overlap=0,
max_num_images=250):
"""
Return array of num windows that can be generated with window_params
that do not overlap with ground truth by more than max_overlap.
* If cls is not given, returns ground truth for all classes.
* If max_num_images is given, samples from at most that many images.
"""
sw = SlidingWindows(self, self)
if not window_params:
window_params = sw.get_default_window_params(cls)
all_windows = []
image_inds = self.get_pos_samples_for_class(cls)
max_num = len(image_inds)
inds = image_inds
if max_num_images:
inds = ut.random_subset(image_inds, max_num_images)
num_per_image = round(1. * num / max_num)
for ind in inds:
image = self.images[ind]
windows = image.get_windows(window_params)
gts = image.get_ground_truth(cls)
for gt in gts.arr:
overlaps = BoundingBox.get_overlap(windows[:, :4], gt[:4])
windows = windows[overlaps <= max_overlap, :]
if windows.shape[0] == 0:
continue
ind_to_take = ut.random_subset_up_to_N(windows.shape[0],
num_per_image)
all_windows.append(
np.hstack((windows[ind_to_take, :],
np.tile(ind, (ind_to_take.shape[0], 1)))))
all_windows = np.concatenate(all_windows, 0)
return all_windows[:num, :]
def get_pos_windows(self, cls=None, window_params=None, min_overlap=0.7):
"""
Return array of all ground truth windows for the class, plus windows
that can be generated with window_params that overlap with it by more
than min_overlap.
* If cls not given, return positive windows for all classes.
* If window_params not given, use default for the class.
* Adjust min_overlap to fetch fewer windows.
"""
sw = SlidingWindows(self, self)
if not window_params:
window_params = sw.get_default_window_params(cls)
overlapping_windows = []
image_inds = self.get_pos_samples_for_class(cls)
times = []
window_nums = []
for i in image_inds:
image = self.images[i]
gts = image.get_ground_truth(cls)
if gts.arr.shape[0] > 0:
overlap_wins = gts.arr[:, :4]
overlap_wins = np.hstack((overlap_wins, np.tile(i, (overlap_wins.shape[0], 1))))
overlapping_windows.append(overlap_wins.astype(int))
windows, time_elapsed = image.get_windows(window_params, with_time=True)
window_nums.append(windows.shape[0])
times.append(time_elapsed)
for gt in gts.arr:
overlaps = BoundingBox.get_overlap(windows[:, :4], gt[:4])
overlap_wins = windows[overlaps >= min_overlap, :]
overlap_wins = np.hstack((overlap_wins, np.tile(i, (overlap_wins.shape[0], 1))))
overlapping_windows.append(overlap_wins.astype(int))
windows = windows[overlaps < min_overlap, :]
overlapping_windows = np.concatenate(overlapping_windows, 0)
print(
"Windows generated per image: %d +/- %.3f, in %.3f +/- %.3f sec"
% (np.mean(window_nums), np.std(window_nums), np.mean(times), np.std(times))
)
return overlapping_windows
def get_windows(self, window_params, with_time=False):
"Return all windows that can be generated with given params."
return SlidingWindows.get_windows(self, None, window_params, with_time)
def main():
parser = argparse.ArgumentParser(description='Execute different functions of our system')
parser.add_argument('mode',
choices=[
'window_stats', 'evaluate_metaparams', 'evaluate_jw',
'evaluate_get_pos_windows', 'train_svm',
'extract_sift','extract_assignments','extract_codebook',
'evaluate_jw_grid', 'final_metaparams',
'assemble_dpm_dets','ctfdet','assemble_ctf_dets'
])
parser.add_argument('--test_dataset', choices=['val','test','train'],
default='test', help='dataset to use for testing. the training dataset \
is automatically inferred (val->train and test->trainval).')
parser.add_argument('--first_n', type=int,
help='only take the first N images in the datasets')
parser.add_argument('--bounds', type=str,
help='the start_time and deadline_time for the ImagePolicy and corresponding evaluation. ex: (1,5)')
parser.add_argument('--name', help='name for this run')
parser.add_argument('--priors', default='random', help= \
"list of choice for the policy for selecting the next action. choose from random, oracle,fixed_order, no_smooth, backoff. ex: --priors=random,oracle,no_smooth")
parser.add_argument('--compare_evals', action='store_true',
default=False, help='plot all the priors modes given on same plot'),
parser.add_argument('--detector', choices=['perfect','perfect_with_noise', 'dpm','ctf'],
default='perfect', help='detector type')
parser.add_argument('--force', action='store_true',
default=False, help='force overwrite')
parser.add_argument('--gist', action='store_true',
default=False, help='use GIST as one of the actions')
parser.add_argument('--clear_tmp', action='store_true',
default=False, help='clear the cached windows folder before running'),
parser.add_argument('--feature_type', choices=['sift','dsift'],
default='dsift', help='use this feature type'),
parser.add_argument('--kernel', choices=['chi2','rbf'],
default='chi2', help='kernel to train svm on'),
args = parser.parse_args()
if args.priors:
args.priors = args.priors.split(',')
if args.bounds:
args.bounds = [float(x) for x in re.findall(r'\d+', args.bounds)]
assert(len(args.bounds)==2)
print(args)
# Load the dataset
dataset = Dataset('full_pascal_'+args.test_dataset)
if args.first_n:
dataset.images = dataset.images[:args.first_n]
# Infer train_dataset
if args.test_dataset=='test':
train_dataset = Dataset('full_pascal_trainval')
elif args.test_dataset=='val':
train_dataset = Dataset('full_pascal_train')
else:
print("Impossible, setting train_dataset to dataset")
train_dataset = dataset
# Create window generator
sw = SlidingWindows(dataset,train_dataset)
if args.clear_tmp:
dirname = config.get_sliding_windows_cached_dir(train_dataset.get_name())
shutil.rmtree(dirname)
dirname = config.get_sliding_windows_cached_dir(dataset.get_name())
shutil.rmtree(dirname)
if args.mode=='assemble_dpm_dets':
policy = DatasetPolicy(dataset,train_dataset,sw)
dets = policy.load_ext_detections(dataset,suffix='dpm_may25')
if args.mode=='assemble_ctf_dets':
policy = DatasetPolicy(dataset,train_dataset,sw)
dets = policy.load_ext_detections(dataset,'ctf','ctf_default')
dets = policy.load_ext_detections(dataset,'ctf','ctf_nohal')
dets = policy.load_ext_detections(dataset,'ctf', 'ctf_halfsize')
if args.mode=='evaluate_get_pos_windows':
evaluate_get_pos_windows(train_dataset)
return
if args.mode=='window_stats':
"Compute and plot the statistics of ground truth window parameters."
results = SlidingWindows.get_dataset_window_stats(train_dataset,plot=True)
if args.mode=='ctfdet':
"""Run Pedersoli's detector on the dataset and assemble into one Table."""
run_pedersoli(dataset)
if args.mode=='evaluate_jw':
"""
Evaluate the jumping window approach by producing plots of recall vs.
#windows.
"""
# TODO hack: both sw and jw should subclass something like WindowGenerator
jw = JumpingWindowsDetector(use_scale=True)
sw.jw = jw
#classes = dataset.classes
classes = ['car']
# classes = ['bicycle' ,'car','horse', 'sofa',\
# 'bird', 'chair', 'motorbike', 'train',\
# 'boat', 'cow', 'person', 'tvmonitor',\
# 'bottle','diningtable', 'pottedplant',\
# 'bus','dog' ,'sheep']
for cls_idx in range(comm_rank, len(classes), comm_size):
#for cls in dataset.classes:
cls = classes[cls_idx]
dirname = config.get_jumping_windows_dir(dataset.get_name())
filename = os.path.join(dirname,'%s'%cls)
sw.evaluate_recall(cls, filename, metaparams=None, mode='jw', plot=True)
if args.mode=='evaluate_jw_grid':
"""
Evaluate the jumping window approach by producing plots of recall vs.
#windows.
"""
sw = SlidingWindows(dataset,train_dataset)
jw = JumpingWindowsDetectorGrid()
sw.jw = jw
for cls in dataset.classes:
dirname = config.get_jumping_windows_dir(dataset.get_name())
filename = os.path.join(dirname,'%s'%cls)
if os.path.isfile(config.data_dir + 'JumpingWindows/'+cls):
sw.evaluate_recall(cls, filename, metaparams=None, mode='jw', plot=True)
if args.mode=='train_svm':
randomize = not os.path.exists('/home/tobibaum')
d = Dataset('full_pascal_train')
dtest = Dataset('full_pascal_val')
e = Extractor()
classes = config.pascal_classes
num_words = 3000
iters = 5
feature_type = 'dsift'
codebook_samples = 15
num_pos = 'max'
testsize = 'max'
if args.first_n:
num_pos = args.first_n
testsize = 1.5*num_pos
kernel = args.kernel
if comm_rank == 0:
ut.makedirs(config.data_dir + 'features/' + feature_type + '/times/')
ut.makedirs(config.data_dir + 'features/' + feature_type + '/codebooks/times/')
ut.makedirs(config.data_dir + 'features/' + feature_type + '/svms/train_times/')
for cls_idx in range(comm_rank, len(classes), comm_size):
#for cls in classes:
cls = classes[cls_idx]
codebook = e.get_codebook(d, feature_type)
pos_arr = d.get_pos_windows(cls)
neg_arr = d.get_neg_windows(pos_arr.shape[0], cls, max_overlap=0)
if not num_pos == 'max':
if not randomize:
pos_arr = pos_arr[:num_pos]
neg_arr = pos_arr[:num_pos]
else:
rand = np.random.random_integers(0, pos_arr.shape[0] - 1, size=num_pos)
pos_arr = pos_arr[rand]
rand = np.random.random_integers(0, neg_arr.shape[0] - 1, size=num_pos)
neg_arr = neg_arr[rand]
pos_table = Table(pos_arr, ['x','y','w','h','img_ind'])
neg_table = Table(neg_arr, pos_table.cols)
train_with_hard_negatives(d, dtest, num_words,codebook_samples,codebook,\
cls, pos_table, neg_table,feature_type, \
iterations=iters, kernel=kernel, L=2, \
testsize=testsize,randomize=randomize)
if args.mode=='evaluate_metaparams':
"""
Grid search over metaparams values for get_windows_new, with the AUC of
recall vs. # windows evaluation.
"""
sw.grid_search_over_metaparams()
return
if args.mode=='final_metaparams':
dirname = config.get_sliding_windows_metaparams_dir(train_dataset.get_name())
# currently these are the best auc/complexity params
best_params_for_classes = [
(62,15,12,'importance',0), #aeroplane
(83,15,12,'importance',0), #bicycle
(62,15,12,'importance',0), #bird
(62,15,12,'importance',0), #boat
(125,12,12,'importance',0), #bottle
(83,12,9,'importance',0), #bus
(125,15,9,'importance',0), #car
(125,12,12,'linear',0), #cat
(125,15,9,'importance',0), #chair
(125,9,6,'importance',0), #cow
(125,15,6,'linear',0), #diningtable
(62,15,12,'importance',0), #dog
(83,15,6,'importance',0), #horse
(83,12,6,'importance',0), #motorbike
(83,15,12,'importance',0), #person
(83,15,6,'importance',0), #pottedplant
(83,15,12,'importance',0), #sheep
(83,9,6,'importance',0), #sofa
(62,12,6,'importance',0), #train
(62,12,12,'importance',0), #tvmonitor
(125,9,12,'importance',0) #all
]
# ACTUALLY THEY ARE ALL THE SAME!
cheap_params = (62, 9, 6, 'importance', 0)
for i in range(comm_rank,dataset.num_classes(),comm_size):
cls = dataset.classes[i]
best_params = best_params_for_classes[i]
#samples,num_scales,num_ratios,mode,priority,cls = cheap_params
metaparams = {
'samples_per_500px': samples,
'num_scales': num_scales,
'num_ratios': num_ratios,
'mode': mode,
'priority': 0 }
filename = '%s_%d_%d_%d_%s_%d'%(
cls,
metaparams['samples_per_500px'],
metaparams['num_scales'],
metaparams['num_ratios'],
metaparams['mode'],
metaparams['priority'])
filename = os.path.join(dirname,filename)
tables = sw.evaluate_recall(cls,filename,metaparams,'sw',plot=True,force=False)
metaparams = {
'samples_per_500px': samples,
'num_scales': num_scales,
'num_ratios': num_ratios,
'mode': mode,
'priority': 1 }
filename = '%s_%d_%d_%d_%s_%d'%(
cls,
metaparams['samples_per_500px'],
metaparams['num_scales'],
metaparams['num_ratios'],
metaparams['mode'],
metaparams['priority'])
filename = os.path.join(dirname,filename)
tables = sw.evaluate_recall(cls,filename,metaparams,'sw',plot=True,force=False)
return
if args.mode=='extract_sift':
e=Extractor()
e.extract_all(['sift'], ['full_pascal_trainval','full_pascal_test'], 0, 0)
if args.mode=='extract_assignments':
e=Extractor()
feature_type = 'sift'
for image_set in ['full_pascal_trainval','full_pascal_test']:
d = Dataset(image_set)
codebook = e.get_codebook(d, feature_type)
print 'codebook loaded'
for img_ind in range(comm_rank,len(d.images),comm_size):
img = d.images[img_ind]
#for img in d.images:
e.get_assignments(np.array([0,0,img.size[0],img.size[1]]), feature_type, \
codebook, img)
if args.mode=='extract_codebook':
d = Dataset('full_pascal_trainval')
e = Extractor()
codebook = e.get_codebook(d, args.feature_type)
def get_windows(self,window_params,with_time=False): "Return all windows that can be generated with given params." return SlidingWindows.get_windows(self,None,window_params,with_time)
