PartImageNet is a paper published in ECCV2022, which proposes an object part-level annotation dataset with the largest number of current categories and images.
PartImageNet: A Large, High-Quality Dataset of Parts
paper: https://arxiv.org/abs/2112.00933
code: https://github.com/TACJu/PartImageNet
The download link of the data set has been released on GitHub and can be downloaded.
The PartImageNet dataset contains 158 categories, with a total of 24095 images. Each image contains component-level annotations of a single object. An example of the annotation is shown in the figure below.
The specific category information of this data set is shown in the table below. 158 subcategories belong to 11 categories. The number of subcategories in each category is given in brackets. The categories are the same. Taking the Fish category as an example, it contains 10 subcategories, and the part categories of objects in each category are Head, Body, Fin, and Tail.
Among the 158 classes, there are 118 non-rigid body classes (such as dogs) and another 40 are rigid body classes (such as cars). In addition, the dataset also provides more fine-grained classification information, as shown in the figure below. Example: Quadruped → Dog → Gordon setter.
For the Semantic Part Segmentation task, the data set is divided according to 85%, 5%, and 10%. The specific information is as follows:
By the way, take a look at the performance indicators of the existing methods on this task, as follows:
For the Few-shot Learning task, the data set is divided according to the number of categories 109, 19, and 30. The specific information is as follows:
