def __init__(self, root_dir, split="train", hd=False):
self.root_dir = root_dir
self.split = split
self.imgs_dir = os.path.join(root_dir, "RGB_Images")
self.masks_dir = os.path.join(root_dir, "Masks")
base_size = 1080 if hd else 513
self.transform_train = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.RandomScaleCrop(base_size=base_size, crop_size=513, fill=0),
tr.RandomRotate(15),
tr.RandomGaussianBlur(),
tr.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
tr.ToTensor()
])
self.transform_validation = transforms.Compose([
tr.FixScaleCrop(crop_size=base_size),
tr.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
tr.ToTensor()
])
with open(os.path.join(root_dir, "Sets", f"{split}.txt"), 'r') as f:
basenames, imgs, masks = [], [], []
for basename in f:
img, mask = self.get_filenames(basename)
basenames.append(basename)
imgs.append(img)
masks.append(mask)
self.files = pd.DataFrame(
list(zip(basenames, imgs, masks)),
columns=['basename', 'img_filename', 'label_filename'])
def __init__(self,
root_dir,
split="train",
train_ratio=0.9,
transform=None,
seed=1234):
self.root_dir = root_dir
self.depths_dir = os.path.join(root_dir, "Depths")
self.images_dir = os.path.join(root_dir, "Images")
self.masks_dir = os.path.join(root_dir, "Masks")
self.overlays_dir = os.path.join(root_dir, "Overlays")
self.plys_dir = os.path.join(root_dir, "PLYs")
self.xmls_dir = os.path.join(root_dir, "XMLs")
self.split = split
self.weights = [1, 1, 1, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1]
self.transform_train = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.RandomScaleCrop(base_size=(287, 352),
crop_size=(287, 352),
fill=0),
tr.RandomRotate(15),
tr.RandomGaussianBlur(),
tr.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
tr.ToTensor()
])
self.transform_validation = transforms.Compose([
tr.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
tr.ToTensor()
])
xml_files = []
for dir_path, dir_names, file_names in os.walk(self.xmls_dir):
xml_files += [os.path.join(dir_path, file) for file in file_names]
sequences = [file.split("/")[-2] for file in xml_files]
frame_ids = [file.split("/")[-1].split(".")[0] for file in xml_files]
self.files = pd.DataFrame(list(zip(sequences, frame_ids)),
columns=['sequence', 'frame_id'])
# self.files = self.files[:100]
np.random.seed(seed)
mask = np.random.rand(len(self.files)) < train_ratio
if split == "train":
self.files = self.files[mask]
else:
self.files = self.files[~mask]
self.files = self.files.reset_index()
def __init__(self, root_dir, split="train"):
self.root_dir = root_dir
self.transform = transforms.Compose([tr.Normalize(mean=(0.5, 0.5, 0.5),
std=(0.5, 0.5, 0.5)),
tr.ToTensor()])
self.images = [f for f in os.listdir(self.root_dir) if os.path.isfile(os.path.join(self.root_dir, f))]
def __init__(self,
root_dir,
split="train",
sampeled_images_path="/home/deepsight/DeepSightData",
train_range=2000,
eval_distance=5):
self.root_dir = root_dir
self.split = split
self.imgs_dir = os.path.join(root_dir, "RGB_Images")
self.masks_dir = os.path.join(root_dir, "Masks")
self.unlabeled_dir = os.path.join(root_dir, "unlabeled_rgb_images")
self.train_range = train_range
self.eval_distance = eval_distance
self.datasets = ['Feb_11_CDE', 'Feb_8_CDE', 'March_1_CDE']
self.mapping = video_scene_to_name(self.datasets, sampeled_images_path)
self.transform_train = transforms.Compose([
tr.RandomHorizontalFlip(temporal=True),
tr.RandomScaleCrop(base_size=513,
crop_size=513,
fill=0,
temporal=True),
tr.RandomRotate(15, temporal=True),
tr.RandomGaussianBlur(temporal=True),
tr.Normalize(mean=(0.5, 0.5, 0.5),
std=(0.5, 0.5, 0.5),
temporal=True),
tr.ToTensor(temporal=True)
])
self.transform_validation = transforms.Compose([
tr.FixScaleCrop(crop_size=513, temporal=True),
tr.Normalize(mean=(0.5, 0.5, 0.5),
std=(0.5, 0.5, 0.5),
temporal=True),
tr.ToTensor(temporal=True)
])
with open(os.path.join(root_dir, "Sets", f"{split}.txt"), 'r') as f:
scenes, ids = [], []
for basename in f:
scene, id = basename.strip("\n").split("scene")
scenes.append(scene)
ids.append(id)
self.files = pd.DataFrame(list(zip(scenes, ids)),
columns=['scene', 'id'])
def transform_val(self, sample):
composed_transforms = transforms.Compose([
tr.FixScaleCrop(crop_size=self.crop_size),
tr.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
tr.ToTensor()])
return composed_transforms(sample)
def transform_tr(self, sample):
composed_transforms = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.RandomRotateB(),
tr.RandomScaleCrop(base_size=self.base_size, crop_size=self.crop_size, fill=255),
tr.RandomGaussianBlur(),
tr.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
tr.ToTensor()])
return composed_transforms(sample)
def __init__(self, root_dir, split="train", transform=None, seed=1234):
self.root_dir = root_dir
self.depths_dir = os.path.join(root_dir, "Depths")
self.images_dir = os.path.join(root_dir, "Images")
self.masks_dir = os.path.join(root_dir, "Masks")
self.overlays_dir = os.path.join(root_dir, "Overlays")
self.plys_dir = os.path.join(root_dir, "PLYs")
self.xmls_dir = os.path.join(root_dir, "XMLs")
self.split = split
# self.weights = [True, True, True, True, True, False, False, True, False, False, False, False, False]
np.random.seed(seed)
# RandomNoise needs to be after RandomBrightnessContrast due to the requirement of uint8 in RandomBrightnessContrast
self.augmentation = aug.Compose([
aug.RandomBrightnessContrast(
p=0.5), # , brightness_limit=0.5, contrast_limit=0.5),
RandomNoise(p=0.5),
])
self.normalize = tr.Normalize(mean=(0.5, 0.5, 0.5),
std=(0.5, 0.5, 0.5),
multiview=True)
self.to_tensor = tr.ToTensor(multiview=True)
# find the sequences
self.sequences = os.listdir(self.xmls_dir)
self.sequences.sort()
# read the frames
self.files = [[], [], [], []]
for idx, seq in enumerate(self.sequences):
self.files[idx] = os.listdir(os.path.join(self.xmls_dir, seq))
self.files[idx].sort()
self.files[idx] = [file.split('.')[0] for file in self.files[idx]]
# read data split index
if self.split == "train":
filename = "train_data.txt"
elif self.split == "validation":
filename = "validation_data.txt"
file = open(os.path.join(self.root_dir, filename), 'r').readlines()
# get indexes
self.index = []
for i in range(len(file)):
self.index.append(int(i))
# LOAD
checkpoint = torch.load(args.resume)
spatial_model.module.load_state_dict(checkpoint['spatial_model_state_dict'])
temporal_model.module.load_state_dict(checkpoint['temporal_model_state_dict'])
#EVAL
temporal_model.eval()
return temporal_model
rgb_transform = transforms.Compose([tr.FixScaleCrop(crop_size=513),
tr.Normalize(mean=(0.5, 0.5, 0.5),
std=(0.5, 0.5, 0.5)),
tr.ToTensor()])
class DeepSightDemoRGB(Dataset):
NUM_CLASSES = 11
CLASSES = ['background', 'ortable', 'psc', 'vsc', 'human', 'cielinglight', 'mayostand', 'table', 'anesthesiacart', 'cannula', 'instrument']
def __init__(self, root_dir):
self.root_dir = root_dir
self.transform = rgb_transform
self.images = [f for f in os.listdir(self.root_dir) if os.path.isfile(os.path.join(self.root_dir, f))]
def __len__(self):
return len(self.images)
def __getitem__(self, item):
path = os.path.join(self.root_dir, self.images[item])