我得到了一个ValueError:所有边界框都应该有正的高度和宽度

Question

嘿，我得到了错误

ValueError: All bounding boxes should have positive height and width. Found invaid box [264.0, 632.0, 264.0, 633.3333740234375] for target at index 2.
Epoch 1/1
Mini-batch: 1/1220 Loss: 0.1509
Mini-batch: 101/1220 Loss: 0.1201
Mini-batch: 201/1220 Loss: 0.1103
Mini-batch: 301/1220 Loss: 0.1098
Mini-batch: 401/1220 Loss: 0.1076
Mini-batch: 501/1220 Loss: 0.1056
Mini-batch: 601/1220 Loss: 0.1044
Mini-batch: 701/1220 Loss: 0.1035
ValueError Traceback (most recent call last)
in ()
13
14 # Calculate losses
—> 15 loss_dict = model(images, targets)
16 batch_loss = sum(loss for loss in loss_dict.values()) / len(loss_dict)
17

1 frames
/usr/local/lib/python3.6/dist-packages/torchvision/models/detection/generalized_rcnn.py in forward(self, images, targets)
91 raise ValueError(“All bounding boxes should have positive height and width.”
92 " Found invaid box {} for target at index {}."
—> 93 .format(degen_bb, target_idx))
94
95 features = self.backbone(images.tensors)

ValueError: All bounding boxes should have positive height and width. Found invaid box [264.0, 632.0, 264.0, 633.3333740234375] for target at index 2.

我在标签csv文件中找不到具有这些坐标的边界框。有谁能帮我解决这个问题吗。

以下是我的数据集类

from torch.utils.data import Dataset, DataLoader

Inherit from pytorch Dataset for convenience
class DamageDataset(Dataset):
def __init__(self, dataframe):

    super().__init__()

    self.filename = dataframe['filename'].unique()

    self.df = dataframe

def __len__(self) -> int:

    return len(self.filename)



def __getitem__(self, index: int):

    filename = self.filename[index]

    image = read_image_from_train_folder(filename).astype(np.float32)

    # Scale to [0,1] range expected by the pre-trained model

    image /= 255.0

    # Convert the shape from [h,w,c] to [c,h,w] as expected by pytorch

    image = torch.from_numpy(image).permute(2,0,1)

    

    records = self.df[self.df['filename'] == filename]

    

    boxes = records[['xmin', 'ymin', 'xmax', 'ymax']].values

    classes= records['class'].values

    damage_labels=[]

    damage_dict={

        'D00': 1,

        'D10': 2,

        'D20': 3,

        'D40': 4,

                                                                    

    }

    for label in classes:

      damage_labels.append(damage_dict[label])

   

    boxes = torch.as_tensor(boxes, dtype=torch.float32)

    n_boxes = boxes.shape[0]

    

    # there is only one foreground class, WHEAT

    labels = torch.as_tensor(damage_labels, dtype=torch.float32)

            

    target = {}

    target['boxes'] = boxes

    target['labels'] = labels

    

    return image, target

这是我的火车代码：

num_epochs = 1

Prepare the model for training
model = model.to(device)

model.train()

for epoch in range(num_epochs):

print("Epoch %i/%i " % (epoch + 1, num_epochs) )

average_loss = 0

for batch_id, (images, targets) in enumerate(train_data_loader):

    # Prepare the batch data

    images, targets = move_batch_to_device(images, targets)

    # Calculate losses

    loss_dict = model(images, targets)

    batch_loss = sum(loss for loss in loss_dict.values()) / len(loss_dict)

    

    # Refresh accumulated optimiser state and minimise losses

    optimizer.zero_grad()

    batch_loss.backward()

有人能帮我找出这个边界框的索引吗?这样我就可以删除它了，我已经用下面的代码遍历了我的数据帧：

for idx, row in merge_labels.iterrows():
if(row[‘xmin’]==264 and row[‘ymin’]== 632 and row[‘xmax’]== 264 and row[‘ymax’]== 633.3333740234375 ):

print(idx)

但是它不打印任何索引。谢谢

Answer 1

这是因为在检测模块的fasterRCNN中应用了调整大小转换。如果您显式地应用调整大小操作，则边界框生成的坐标将根据调整大小定义进行更改，但如果您尚未应用调整大小转换，并且图像的最小和最大大小为outsider (800,1333)，则会应用默认的调整大小转换。

查看pytorch git repo中的以下代码片段。

generalized_rcnn.py :/torchvision/models/detection，在Module中产生异常

        degenerate_boxes = boxes[:, 2:] <= boxes[:, :2]
        if degenerate_boxes.any():
            # print the first degenerate box
            bb_idx = torch.where(degenerate_boxes.any(dim=1))[0][0]
            degen_bb: List[float] = boxes[bb_idx].tolist()
            raise ValueError("All bounding boxes should have positive height and width."
                             " Found invalid box {} for target at index {}."
                             .format(degen_bb, target_idx))

默认最小最大图像大小；模块:检测/ faster_rcnn

def __init__(self, backbone, num_classes=None,
             # transform parameters
             min_size=800, max_size=1333,
             image_mean=None, image_std=None,
             ............................

transform = GeneralizedRCNNTransform(min_size, max_size, image_mean, image_std)

调整目标大小；模块:检测/ transform.py

def resize_boxes(boxes, original_size, new_size):
    # type: (Tensor, List[int], List[int]) -> Tensor
    ratios = [
        torch.tensor(s, dtype=torch.float32, device=boxes.device) /
        torch.tensor(s_orig, dtype=torch.float32, device=boxes.device)
        for s, s_orig in zip(new_size, original_size)
    ]
    ratio_height, ratio_width = ratios
    xmin, ymin, xmax, ymax = boxes.unbind(1)

    xmin = xmin * ratio_width
    xmax = xmax * ratio_width
    ymin = ymin * ratio_height
    ymax = ymax * ratio_height
    return torch.stack((xmin, ymin, xmax, ymax), dim=1)

您可以通过删除/更正原始数据集中xmin和xmax或ymin和ymax相等的边界框来修复此问题。

Answer 2

我认为你需要检查你的图像边界框。我通过确保我的xmin < xmax和ymin < ymax解决了这个问题。如果xmin等于xmax，那么它将是一行