ccvl/ImageNet3D

---
license: cc-by-nc-4.0
pretty_name: imagenet3d
---
## ImageNet3D

We present **ImageNet3D**, a large dataset for general-purpose object-level 3D understanding. ImageNet3D augments 200 categories from the ImageNet dataset with 2D bounding box, 3D pose, 3D location annotations, and image captions interleaved with 3D information.

Refer to [github.com/wufeim/imagenet3d](https://github.com/wufeim/imagenet3d) for the full documentation and sample preprocessing code for ImageNet3D.

### Download Data

**ImageNet3D-v1.0:** Directly download from the HuggingFace WebUI, or on a server, run

```sh
wget https://huggingface.co/datasets/ccvl/ImageNet3D/resolve/main/imagenet3d_v1.zip
```

**Future updates:** We are working on annotating more object categories and improving the quality of current annotations.
The next update is planned to be released by the end of Jan 2025. Please let us know if you have any suggestions for future updates.

### Example Usage

```py
from PIL import Image
import numpy as np

img_path = 'imagenet3d/bed/n02818832_13.JPEG'
annot_path = 'imagenet3d/bed/n02818832_13.npz'

img = np.array(Image.open(img_path).convert('RGB'))
annot = dict(np.load(annot_path, allow_pickle=True))['annotations']

# Number of objects
num_objects = len(annot)

# Annotation of the first object
azimuth = annot[0]['azimuth'] # float, [0, 2*pi]
elevation = annot[0]['elevation'] # float, [0, 2*pi]
theta = annot[0]['theta'] # float, [0, 2*pi]
cad_index = annot[0]['cad_index'] # int
distance = annot[0]['distance'] # float
viewport = annot[0]['viewport'] # int
img_height = annot[0]['height'] # numpy.uint16
img_width = annot[0]['width'] # numpy.uint16
bbox = annot[0]['bbox'] # numpy.ndarray, (x1, y1, x2, y2)
category = annot[0]['class'] # str
principal_x = annot[0]['px'] # float
principal_y = annot[0]['py'] # float

# label indicating the quality of the object, occluded or low quality
object_status = annot[0]['object_status'] # str, one of ('status_good', 'status_partially', 'status_barely', 'status_bad')

# label indicating if multiple objects from same category very close to each other
dense = annot[0]['dense'] # str, one of ('dense_yes', 'dense_no')
```

---
license: CC BY-NC 4.0（知识共享署名-非商业性使用4.0国际许可协议）
pretty_name: ImageNet3D
---
## ImageNet3D

我们推出**ImageNet3D**，一款面向通用目标级三维理解的大型数据集。ImageNet3D 对 ImageNet 数据集中的200个类别进行扩展，为其添加了二维边界框（2D bounding box）、三维姿态（3D pose）、三维位置（3D location）标注，以及融合三维信息的图像标题。

如需获取 ImageNet3D 的完整文档与示例预处理代码，请参考 [github.com/wufeim/imagenet3d](https://github.com/wufeim/imagenet3d)。

### 数据下载

**ImageNet3D-v1.0**：可直接通过 HuggingFace 网页界面下载；若在服务器环境中，可执行以下命令：
sh
wget https://huggingface.co/datasets/ccvl/ImageNet3D/resolve/main/imagenet3d_v1.zip

**后续更新计划**：我们正致力于为更多物体类别添加标注，并优化现有标注的质量。下一版更新预计于2025年1月底前发布。若您对后续更新有任何建议，欢迎随时告知。

### 示例用法

py
from PIL import Image
import numpy as np

img_path = 'imagenet3d/bed/n02818832_13.JPEG'
annot_path = 'imagenet3d/bed/n02818832_13.npz'

img = np.array(Image.open(img_path).convert('RGB'))
annot = dict(np.load(annot_path, allow_pickle=True))['annotations']

# 物体总数
num_objects = len(annot)

# 首个物体的标注信息
# 方位角（azimuth），浮点型，取值范围为[0, 2π]
azimuth = annot[0]['azimuth']
# 仰角（elevation），浮点型，取值范围为[0, 2π]
elevation = annot[0]['elevation']
# 旋转角（theta），浮点型，取值范围为[0, 2π]
theta = annot[0]['theta']
# CAD模型索引（cad_index），整型
cad_index = annot[0]['cad_index']
# 观测距离（distance），浮点型
distance = annot[0]['distance']
# 视口编号（viewport），整型
viewport = annot[0]['viewport']
# 图像高度（height），numpy.uint16 类型
img_height = annot[0]['height']
# 图像宽度（width），numpy.uint16 类型
img_width = annot[0]['width']
# 二维边界框（bbox），numpy.ndarray 类型，格式为 (x1, y1, x2, y2)
bbox = annot[0]['bbox']
# 类别标签（class），字符串类型
category = annot[0]['class']
# 主点x坐标（px），浮点型
principal_x = annot[0]['px']
# 主点y坐标（py），浮点型
principal_y = annot[0]['py']

# 物体质量标签，用于标识物体是否被遮挡或画质不佳
# 可选值为 ('status_good', 'status_partially', 'status_barely', 'status_bad')
object_status = annot[0]['object_status']

# 密集度标签，用于标识同类物体是否彼此距离过近
# 可选值为 ('dense_yes', 'dense_no')
dense = annot[0]['dense']