def __init__(self, do_cuda=True):
print('Using TfeatFeature2D')
self.model_base_path = config.cfg.root_folder + '/thirdparty/tfeat/'
self.do_cuda = do_cuda & torch.cuda.is_available()
print('cuda:', self.do_cuda)
device = torch.device("cuda:0" if self.do_cuda else "cpu")
torch.set_grad_enabled(False)
# mag_factor is how many times the original keypoint scale
# is enlarged to generate a patch from a keypoint
self.mag_factor = 3
print('==> Loading pre-trained network.')
#init tfeat and load the trained weights
self.model = tfeat_model.TNet()
self.models_path = self.model_base_path + 'pretrained-models'
self.net_name = 'tfeat-liberty'
self.model.load_state_dict(
torch.load(
os.path.join(self.models_path, self.net_name + ".params")))
if self.do_cuda:
self.model.cuda()
print('Extracting on GPU')
else:
print('Extracting on CPU')
self.model = model.cpu()
self.model.eval()
print('==> Successfully loaded pre-trained network.')
import os
import cv2
import tfeat_utils
import numpy as np
import cv2
from matplotlib import pyplot as plt
tfeat_base_path = '../../thirdparty/tfeat/'
# open images
img1 = cv2.imread(tfeat_base_path + 'imgs/v_churchill/1.ppm', 0)
img2 = cv2.imread(tfeat_base_path + 'imgs/v_churchill/6.ppm', 0)
#init tfeat and load the trained weights
tfeat = tfeat_model.TNet()
models_path = tfeat_base_path + 'pretrained-models'
net_name = 'tfeat-liberty'
tfeat.load_state_dict(
torch.load(os.path.join(models_path, net_name + ".params")))
tfeat.cuda()
tfeat.eval()
# use BRISK detector
brisk = cv2.BRISK_create()
kp1, des1 = brisk.detectAndCompute(img1, None)
kp2, des2 = brisk.detectAndCompute(img2, None)
bf = cv2.BFMatcher(cv2.NORM_HAMMING)
matches = bf.knnMatch(des1, des2, k=2)
# Apply ratio test
def main():
args = parse_args()
tfeat = tfeat_model.TNet()
tfeat.load_state_dict(torch.load(args.model_path))
tfeat.cuda()
tfeat.eval()
if not os.path.exists(os.path.join(args.dataset_path, "descriptors")):
os.makedirs(os.path.join(args.dataset_path, "descriptors"))
image_names = os.listdir(os.path.join(args.dataset_path, "images"))
for i, image_name in enumerate(image_names):
patches_path = os.path.join(args.dataset_path, "descriptors",
image_name + ".bin.patches.mat")
if not os.path.exists(patches_path):
continue
print("Computing features for {} [{}/{}]".format(
image_name, i + 1, len(image_names)),
end="")
start_time = time.time()
descriptors_path = os.path.join(args.dataset_path, "descriptors",
image_name + ".bin")
if os.path.exists(descriptors_path):
print(" -> skipping, already exist")
continue
with h5py.File(patches_path, 'r') as patches_file:
patches31 = np.array(patches_file["patches"]).T
if patches31.ndim != 3:
print(" -> skipping, invalid input")
write_matrix(descriptors_path, np.zeros((0, 128),
dtype=np.float32))
continue
patches = np.empty((patches31.shape[0], 32, 32), dtype=np.float32)
patches[:, :31, :31] = patches31
patches[:, 31, :31] = patches31[:, 30, :]
patches[:, :31, 31] = patches31[:, :, 30]
patches[:, 31, 31] = patches31[:, 30, 30]
descriptors = []
for i in range(0, patches.shape[0], args.batch_size):
patches_batch = \
patches[i:min(i + args.batch_size, patches.shape[0])]
patches_batch = \
torch.from_numpy(patches_batch[:, None]).float().cuda()
descriptors.append(tfeat(patches_batch).detach().cpu().numpy())
if len(descriptors) == 0:
descriptors = np.zeros((0, 128), dtype=np.float32)
else:
descriptors = np.concatenate(descriptors)
write_matrix(descriptors_path, descriptors)
print(" in {:.3f}s".format(time.time() - start_time))