README.md

---
license: mit
task_categories:
- image-classification
- image-segmentation
size_categories:
- 100K<n<1M
pretty_name: PartImageNet++
viewer: false
---

## PartImageNet++ Dataset

[Paper](https://huggingface.co/papers/2601.01454) | [GitHub](https://github.com/LixiaoTHU/PartImageNetPP)

PartImageNet++ (PIN++) is an extensive dataset designed for robust object recognition and segmentation tasks. This dataset expands upon the original ImageNet dataset by providing detailed part annotations for each object category.
Official repository for [PartImageNet++ Dataset: Scaling up Part-based Models for Robust Recognition](https://huggingface.co/papers/2601.01454).

### Dataset Statistics

The dataset includes:
- **1000 object categories** derived from the original ImageNet-1K.
- **3308 part categories** representing different parts of objects.
- **100,000 annotated images**, with each object category containing 100 images (downloaded from the ImageNet-1K dataset).
- **406,364 part mask annotations** ensuring comprehensive coverage and detailed segmentation.

### Usage

Train a model, e.g., an MPM with ResNet-50-GELU as the backbone by AT:

```bash
python -m torch.distributed.launch --nproc_per_node=8 adv_train.py --configs=./train_configs/1k_pp/resnet50_part_pp_init.yaml

python -m torch.distributed.launch --nproc_per_node=8 adv_train.py --configs=./train_configs/1k_pp/resnet50_part_pp_at.yaml
```

Evaluate the trained models with Autoattack under different attack threats, e.g., the MPM model under L2 attack with $\epsilon = 2$ :

```bash
python eval.py \
    --configs=./train_configs/1k_pp/resnet50_part_pp_at.yaml \
    --checkpoint=/home/user/ssd/*/checkpoint.pth \
    --attack_types="autoattack" \
    --norm "L2" \
    --eval-eps 2
```

### Structure and Contents

Each JSON file in the `json` directory represents one object category and its corresponding part annotations. 

The `including` folder provides detailed inclusion relations of parts, illustrating hierarchical relationships between different part categories.

The `discarded_data.json` file lists low-quality images excluded from the dataset to maintain high annotation standards.

The `category_name.json` file contains each JSON file's file name, along with its corresponding part name and object name.

### Visualizations

We provide a visualization demo tool to explore and inspect the annotations. This tool helps users better understand the dataset's structure and details.

### Citation
If you find this useful in your research, please cite this work:
```bibtex
@inproceedings{li2024pinpp,
  author = {Li, Xiao and Liu, Yining and Dong, Na and Qin, Sitian and Hu, Xiaolin},
  title = {PartImageNet++ Dataset: Scaling up Part-based Models for Robust Recognition},
  booktitle={European conference on computer vision},
  year = {2024},
  organization={Springer}
}
```