Prisma-Multimodal/segmented-imagenet1k-subset

Segmented ImageNet-1K Subset是ImageNet-1K的一个子集，包含了实例分割的注释（类别、边界框和掩码）。这些注释是由Grounded Segment Anything自动生成的。数据集总共有12,000张图像，其中10,000张来自ImageNet-1K的训练集，1,000张分别来自测试集和验证集。图像和掩码的组织结构与ImageNet-1K相同，并且提供了用于加载数据的代码示例。

原始信息汇总

Segmented ImageNet-1K Subset

概述

• 数据集名称: Segmented ImageNet-1K Subset
• 数据集大小: 12,000 张图片
  • 10,000 张来自 ImageNet-1K 训练集
  • 1,000 张来自测试集
  • 1,000 张来自验证集
• 注释类型: 实例分割注释（类别、边界框和掩码）
• 注释生成工具: Grounded Segment Anything

数据组织

• 图片组织结构:

  images/ train_images/ val_images/ test_images/

• 掩码组织结构:

  masks/ train_masks/ val_masks/ test_masks/

• 注释文件: train.json, val.json, test.json

  • 包含图片路径、分数、边界框、标签和掩码路径信息

示例注释文件内容

json { "image": "images/val_images/ILSVRC2012_val_00000025_n01616318.JPEG", "scores": [0.5, 0.44, 0.43, 0.28], "boxes": [[149, 117, 400, 347], [2, 2, 498, 497], [148, 115, 401, 349], [2, 2, 498, 497]], "labels": ["bird", "dirt field", "vulture", "land"], "masks": ["masks/val_masks/ILSVRC2012_val_00000025_n01616318_00.png", "masks/val_masks/ILSVRC2012_val_00000025_n01616318_01.png", "masks/val_masks/ILSVRC2012_val_00000025_n01616318_02.png", "masks/val_masks/ILSVRC2012_val_00000025_n01616318_03.png"] }

数据加载器示例

python class PatchDataset(Dataset): def init(self, dataset, patch_size=16, width=224, height=224): self.dataset = dataset self.transform = transforms.Compose([ transforms.Resize((width, height)), transforms.ToTensor(), ]) self.patch_size = patch_size self.width = width self.height = height

def __len__(self):
    return len(self.dataset)

def __getitem__(self, idx):
    item = self.dataset[idx]
    image = self.transform(item[image])
    masks = item[masks]
    labels = item[labels]
    
    num_patches = self.width // self.patch_size
    label_array = [[[] for _ in range(num_patches)] for _ in range(num_patches)]
    
    for mask, label in zip(masks, labels):
        mask = mask.resize((self.width, self.height))
        mask_array = np.array(mask) > 0
        reduced_mask = self.reduce_mask(mask_array)
        
        for i in range(num_patches):
            for j in range(num_patches):
                if reduced_mask[i, j]:
                    label_array[i][j].append(label)
    
    return image, label_array

def reduce_mask(self, mask):
    new_h = mask.shape[0] // self.patch_size
    new_w = mask.shape[1] // self.patch_size
    
    reduced_mask = np.zeros((new_h, new_w), dtype=bool)
    
    for i in range(new_h):
        for j in range(new_w):
            patch = mask[i*self.patch_size:(i+1)*self.patch_size, j*self.patch_size:(j+1)*self.patch_size]
            reduced_mask[i, j] = np.any(patch)
    
    return reduced_mask

引用

bibtex @misc{segmented_imagenet1k_subset_2024, author = {ViT-Prisma Contributors}, title = {Segmented ImageNet-1k Subset}, url = {https://huggingface.co/datasets/Prisma-Multimodal/segmented-imagenet1k-subset}, version = {1.0.0}, date = {2024-04-02}, }