3
detector.py is an out-of-the-box windowed detector
4
callable from the command line.
6
By default it configures and runs the Caffe reference ImageNet model.
7
Note that this model was trained for image classification and not detection,
8
and finetuning for detection can be expected to improve results.
10
The selective_search_ijcv_with_python code required for the selective search
11
proposal mode is available at
12
https://github.com/sergeyk/selective_search_ijcv_with_python
15
- batch up image filenames as well: don't want to load all of them into memory
16
- come up with a batching scheme that preserved order / keeps a unique ID
26
CROP_MODES = ['list', 'selective_search']
27
COORD_COLS = ['ymin', 'xmin', 'ymax', 'xmax']
31
pycaffe_dir = os.path.dirname(__file__)
33
parser = argparse.ArgumentParser()
37
help="Input txt/csv filename. If .txt, must be list of filenames.\
38
If .csv, must be comma-separated file with header\
39
'filename, xmin, ymin, xmax, ymax'"
43
help="Output h5/csv filename. Format depends on extension."
48
default=os.path.join(pycaffe_dir,
49
"../models/bvlc_reference_caffenet/deploy.prototxt"),
50
help="Model definition file."
54
default=os.path.join(pycaffe_dir,
55
"../models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel"),
56
help="Trained model weights file."
60
default="selective_search",
62
help="How to generate windows for detection."
67
help="Switch for gpu computation."
71
default=os.path.join(pycaffe_dir,
72
'caffe/imagenet/ilsvrc_2012_mean.npy'),
73
help="Data set image mean of H x W x K dimensions (numpy array). " +
74
"Set to '' for no mean subtraction."
79
help="Multiply input features by this scale to finish preprocessing."
85
help="Multiply raw input by this scale before preprocessing."
90
help="Order to permute input channels. The default converts " +
91
"RGB -> BGR since BGR is the Caffe default by way of OpenCV."
98
help="Amount of surrounding context to collect in input window."
100
args = parser.parse_args()
102
mean, channel_swap = None, None
104
mean = np.load(args.mean_file)
105
if mean.shape[1:] != (1, 1):
106
mean = mean.mean(1).mean(1)
107
if args.channel_swap:
108
channel_swap = [int(s) for s in args.channel_swap.split(',')]
118
detector = caffe.Detector(args.model_def, args.pretrained_model, mean=mean,
119
input_scale=args.input_scale, raw_scale=args.raw_scale,
120
channel_swap=channel_swap,
121
context_pad=args.context_pad)
125
print("Loading input...")
126
if args.input_file.lower().endswith('txt'):
127
with open(args.input_file) as f:
128
inputs = [_.strip() for _ in f.readlines()]
129
elif args.input_file.lower().endswith('csv'):
130
inputs = pd.read_csv(args.input_file, sep=',', dtype={'filename': str})
131
inputs.set_index('filename', inplace=True)
133
raise Exception("Unknown input file type: not in txt or csv.")
136
if args.crop_mode == 'list':
139
(ix, inputs.iloc[np.where(inputs.index == ix)][COORD_COLS].values)
140
for ix in inputs.index.unique()
142
detections = detector.detect_windows(images_windows)
144
detections = detector.detect_selective_search(inputs)
145
print("Processed {} windows in {:.3f} s.".format(len(detections),
149
df = pd.DataFrame(detections)
150
df.set_index('filename', inplace=True)
151
df[COORD_COLS] = pd.DataFrame(
152
data=np.vstack(df['window']), index=df.index, columns=COORD_COLS)
157
if args.output_file.lower().endswith('csv'):
160
class_cols = ['class{}'.format(x) for x in range(NUM_OUTPUT)]
161
df[class_cols] = pd.DataFrame(
162
data=np.vstack(df['feat']), index=df.index, columns=class_cols)
163
df.to_csv(args.output_file, cols=COORD_COLS + class_cols)
166
df.to_hdf(args.output_file, 'df', mode='w')
167
print("Saved to {} in {:.3f} s.".format(args.output_file,
171
if __name__ == "__main__":