R-YOLOv4
R-YOLOv4 copied to clipboard
Published
20 hours ago •
yingkunwu
yingkunwu
数据集格式问题
大佬，请问您一下，你的数据集是标签txt中的格式，和r3det中txt的不太一样，您用的是哪个版本的rolabelimg软件呢，我转换后的txt和您的不太一样，谢谢
不好意思晚了幾天回你。我的格式沒有參考過r3det的，想問你是參考哪邊的格式呢？
我是用labelImg2來幫助我Label的。Label完的格式會跟UCAS的不太一樣，不過只需要從不同格式中，取出bounding box的 （x, y, w, h, angle），分別是中心點、寬、長和角度就可以進行訓練。你可以在load.py裡面做修改。
下方是我針對使用labelImg2做的Label產生的格式做的修改供你參考
class ListDataset(Dataset):
def __init__(self, list_path, img_size=416, augment=True, multiscale=True, normalized_labels=False):
self.img_files = list_path
self.label_files = [
path.replace("images", "labels").replace(".png", ".txt").replace(".jpg", ".txt")
for path in self.img_files
]
self.img_size = img_size
#self.labels = labels
self.max_objects = 100
self.augment = augment
self.multiscale = multiscale
self.normalized_labels = normalized_labels
self.min_size = self.img_size - 3 * 32
self.max_size = self.img_size + 3 * 32
self.batch_count = 0
def __getitem__(self, index):
# ---------
# Image
# ---------
img_path = self.img_files[index]
# Extract image as PyTorch tensor
img = transforms.ToTensor()(Image.open(img_path).convert('RGB'))
# Handle images with less than three channels
if len(img.shape) != 3:
img = img.unsqueeze(0)
img = img.expand((3, img.shape[1:]))
_, h, w = img.shape
h_factor, w_factor = (h, w) if self.normalized_labels else (1, 1)
# Pad to square resolution
if self.augment:
if np.random.random() < 0.25:
img = gaussian_noise(img, 0.0, np.random.random())
if np.random.random() < 0.25:
img = hsv(img)
img, pad = pad_to_square(img, 0)
_, padded_h, padded_w = img.shape
# ---------
# Label
# ---------
label_path = self.label_files[index % len(self.img_files)].rstrip()
if os.path.exists(label_path):
boxes = torch.from_numpy(np.loadtxt(label_path).reshape(-1, 6))
num_targets = len(boxes)
x, y, w, h, theta, label = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3], boxes[:, 4], boxes[:, 5]
temp_theta = []
for t in theta:
if t > np.pi / 2:
t = t - np.pi
elif t <= -(np.pi / 2):
t = t + np.pi
temp_theta.append(t)
theta = torch.stack(temp_theta)
assert (-np.pi / 2 < theta).all() or (theta <= np.pi / 2).all()
for i in range(num_targets):
if w[i] < h[i]:
temp1, temp2 = h[i].clone(), w[i].clone()
w[i], h[i] = temp1, temp2
if theta[i] > 0:
theta[i] = theta[i] - np.pi / 2
else:
theta[i] = theta[i] + np.pi / 2
assert (-np.pi / 2 < theta).all() or (theta <= np.pi / 2).all()
# Extract coordinates for unpadded + unscaled image
x1 = w_factor * (x - w / 2)
y1 = h_factor * (y - h / 2)
x2 = w_factor * (x + w / 2)
y2 = h_factor * (y + h / 2)
# Adjust for added padding
x1 += pad[0]
y1 += pad[2]
x2 += pad[1]
y2 += pad[3]
# Returns (x, y, w, h)
x = ((x1 + x2) / 2) / padded_w
y = ((y1 + y2) / 2) / padded_h
w *= w_factor / padded_w
h *= h_factor / padded_h
targets = torch.zeros((len(boxes), 7))
targets[:, 1] = label
targets[:, 2] = x
targets[:, 3] = y
targets[:, 4] = w
targets[:, 5] = h
targets[:, 6] = theta
assert (0 <= x).all() or (x <= 1).all()
assert (0 <= y).all() or (y <= 1).all()
assert (0 <= w).all() or (w <= 1).all()
assert (0 <= h).all() or (h <= 1).all()
else:
assert False
targets = torch.zeros((1, 7))
targets[:, 1] = -1
return img_path, img, targets
# Apply augmentations
if self.augment:
if np.random.random() < 0.25:
img, targets = rotate(img, targets)
if np.random.random() < 0.5:
img, targets = horisontal_flip(img, targets)
if np.random.random() < 0.5:
img, targets = vertical_flip(img, targets)
return img_path, img, targets
def collate_fn(self, batch):
paths, imgs, targets = list(zip(*batch))
# Remove empty placeholder targets
targets = [boxes for boxes in targets if boxes is not None]
# Add sample index to targets
for i, boxes in enumerate(targets):
boxes[:, 0] = i
targets = torch.cat(targets, 0)
# Selects new image size every tenth batch
if self.multiscale and self.batch_count % 10 == 0:
self.img_size = random.choice(range(self.min_size, self.max_size + 1, 32))
# Resize images to input shape
imgs = torch.stack([resize(img, self.img_size) for img in imgs])
self.batch_count += 1
return paths, imgs, targets
def __len__(self):
return len(self.img_files)
def split_data(data_dir, img_size, batch_size=1, shuffle=True, augment=True, multiscale=False):
files = sorted(glob.glob(data_dir + "/*.jpg"))
train_dataset = ListDataset(files, img_size=img_size, augment=augment, multiscale=multiscale)
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=shuffle,
pin_memory=True, collate_fn=train_dataset.collate_fn)
return train_dataset, train_dataloader
可以参考这个作者的链接 https://github.com/ChenCongGit/RoLabelImg_Transform
大佬可以给个联系方式嘛？还有问题想请教一下
可以参考这个作者的链接 https://github.com/ChenCongGit/RoLabelImg_Transform
大佬可以给个联系方式嘛？还有问题想请教一下
我也没跑通呢，我可以加你一起研究
可以参考这个作者的链接 https://github.com/ChenCongGit/RoLabelImg_Transform
大佬可以给个联系方式嘛？还有问题想请教一下
我也没跑通呢，我可以加你一起研究
qq：171536269