DOR-Bench

DOR-Bench is an image benchmark for evaluating object-removal methods in dense scenes, introduced in the paper DORS: Dynamic Attention Routing for Diffusion-based Object Removal in Dense Scenes.

• Project page: https://httang1224.github.io/DORS/
• Code: https://github.com/httang1224/DORS
• Paper: https://huggingface.co/papers/2607.16656

This release contains 400 test cases. Each case consists of an input image and a binary mask that identifies the object region to be removed. The benchmark focuses on dense scenes containing visually similar instances, where object-removal methods may incompletely erase the target or introduce duplicated structures and residual artifacts.

Dataset structure

.
├── images/             # 400 input images (PNG)
├── masks/              # 400 binary masks (PNG)
├── visualizations/     # 400 preview images (JPEG)
├── metadata.jsonl      # File correspondence and sample IDs
└── README.md

Files are paired by a shared sample ID:

images/dor_001.png
masks/dor_001.png
visualizations/dor_001.jpg

The complete ID range is dor_001 through dor_400.

Data fields

Each row in metadata.jsonl contains:

• id: unique sample identifier.
• image: path to the input image.
• mask: path to the corresponding binary object mask.
• visualization: path to a three-panel preview image.

Masks contain two pixel values: 0 for the background and 255 for the target region.

Each visualization is arranged from left to right as:

input image | binary mask | mask overlaid on the input image

The overlay uses a light-green fill and a green boundary for visual inspection only. Visualizations are not required when evaluating a method.

Intended use

DOR-Bench is intended for evaluating whether an object-removal method completely erases the target specified by a binary mask while producing visually coherent content in the removed region.

The benchmark is provided as a single test split. It is not intended to be used as training data unless explicitly permitted by the applicable data licenses.

Loading with Hugging Face Datasets

After replacing the repository ID below with the published Hugging Face dataset ID:

from datasets import load_dataset

dataset = load_dataset("qc1752/DOR-Bench", split="test")
sample = dataset[0]

image = sample["image"]
mask = sample["mask"]
visualization = sample["visualization"]

Citation

If you find this work useful for your research, please consider citing:

@inproceedings{tang2026dors,
  title     = {{DORS}: Dynamic Attention Routing for Diffusion-based
               Object Removal in Dense Scenes},
  author    = {Tang, Haitong and Liu, Haipeng and Wang, Yang},
  booktitle = {Proceedings of the 34th ACM International Conference
               on Multimedia},
  year      = {2026}
}