def prepare_data():
#Set the training configuration first
cfg_path="models/pvanet/lite/train_1K.yml"
cfg_from_file(cfg_path)
"""
1. PREPARING DATASET
"""
#Firstly, prepare the dataset for fine-tuning
#Different kind of dataset is wrapped by the IMDB class, originally designed by Ross Girshick
#You need to put coco data directory(soft-link works as well) under the PVA-NET directory
#COCO IMDB needs two parameter: data-split and year , "Sedans_1", "Sedans_2"
#coco_train = coco("train", "20SNAPSHOT_ITERS: 6014")
main_classes = CLASS_SETS["3-car"]
mapper = { "trailer-head": '__background__', 'person': '__background__', 'scooter': '__background__', 'bike':'__background__', "motorcycle": "__background__", "bicycle": "__background__", "truck":"__background__", "bus":"__background__"}
vatic_names = ["A1HighwayDay", 'B2HighwayNight', "pickup", "tanktruck", "van", "PU_Van", "Sedans_1", "Sedans_2"]
vatics = [VaticData(vatic_name, main_classes, CLS_mapper=mapper, train_split="all") for vatic_name in vatic_names]
NCTU_VIDEOS = [13, 17, 18, 19, 20, 3, 36, 38, 4, 5 ,6, 7, 8, 9, 10, 11, 12]
NCTU_vatic_names = ["NCTU_{}.MOV".format(video) for video in NCTU_VIDEOS]
NCTU_vatics = [VaticData(vatic_name, main_classes, CLS_mapper=mapper, train_split="all") for vatic_name in NCTU_vatic_names]
imdb_group = IMDBGroup(vatics + NCTU_vatics)
#imdb_group = IMDBGroup(vatics)
imdb, roidb = combined_roidb(imdb_group)
total_len = float(len(imdb_group.gt_roidb()))
print(total_len)
return roidb
def prepare_data():
#Set the training configuration first
cfg_path = "models/pvanet/lite/train.yml"
cfg_from_file(cfg_path)
"""
1. PREPARING DATASET
"""
#Firstly, prepare the dataset for fine-tuning
#Different kind of dataset is wrapped by the IMDB class, originally designed by Ross Girshick
#You need to put coco data directory(soft-link works as well) under the PVA-NET directory
#COCO IMDB needs two parameter: data-split and year
coco_train = coco("train", "2014")
coco_val = coco("val", "2014")
#Fetch the classes of coco dataet, this will be useful in the following section
classes = coco_val._classes
#Next, we import the VOC dataset via pascal_voc wrapper
#Since VOC and COCO data have different naming among classes, a naming mapper is needed to unify the class names
mapper = {
"tvmonitor": "tv",
"sofa": "couch",
"aeroplane": "airplane",
"motorbike": "motorcycle",
"diningtable": "dining table",
"pottedplant": "potted plant"
}
#Finnaly, let's wrap datasets from Vatic.
#A vatic dataset directory should be located under ~/data/ directory in the naming of data-*
#For example: ~/data/data-YuDa, ~/data/data-A1HighwayDay
vatic_names = [
"YuDa", "A1HighwayDay", "B2HighwayNight", "airport", "airport2"
]
mapper = {"van":"car", "trailer-head":"truck",\
"sedan/suv":"car", "scooter":"motorcycle", "bike":"bicycle"}
vatics = [
VaticData(vatic_name, classes, CLS_mapper=mapper, train_split="all")
for vatic_name in vatic_names
]
#Combine all the IMDBs into one single IMDB for training
datasets = vatics + [coco_train, coco_val]
imdb_group = IMDBGroup(datasets)
imdb, roidb = combined_roidb(imdb_group)
total_len = float(len(imdb_group.gt_roidb()))
#Show the dataset percentage in the whole composition
for dataset in imdb_group._datasets:
img_nums = len(dataset.gt_roidb())
print(dataset.name, img_nums,
"{0:.2f}%".format(img_nums / total_len * 100))
return roidb