Transparent-460

Transparent object matting dataset introduced in TransMatting: Enhancing Transparent Objects Matting with Transformers (ECCV 2022).

Contains 460 transparent foreground images with corresponding alpha mattes. Foregrounds are composited onto background images to generate training/test pairs.

Dataset Preview

Dataset Structure

Transparent-460/
├── Train/
│   ├── fg/                              # 410 foreground images (transparent objects)
│   ├── alpha/                           # 410 alpha mattes
│   ├── Composition_code.py             # compositing script
│   ├── transparent-460-train-fg-names.txt
│   └── transparent-460-train-bg-names.txt   # 41,000 COCO BG filenames (100 per FG)
└── Test/
    ├── fg/                              # 50 foreground images
    ├── alpha/                           # 50 alpha mattes
    ├── trimap/                          # 50 trimap masks
    ├── Composition_code.py             # compositing script
    ├── metric_evaluation.py
    ├── transparent-460-test-fg-names.txt
    └── transparent-460-test-bg-names.txt    # 1,000 Pascal VOC BG filenames (20 per FG)

Splits

Data Fields

• fg — foreground image of a transparent object (JPEG)
• alpha — corresponding alpha matte (PNG, grayscale 0–255)
• trimap — ternary region map for test set (PNG; 0=background, 128=unknown, 255=foreground)

Compositing

Background images are not included (COCO / Pascal VOC must be downloaded separately). Use the provided Composition_code.py to composite FG images onto BG images:

# Each train FG is composited onto 100 COCO BG images.
# Each test FG is composited onto 20 Pascal VOC BG images.
python Composition_code.py

Example Usage

from datasets import load_dataset

ds = load_dataset("Thinnaphat/transparent-460")

# Train split
train_sample = ds["train"][0]
fg_image  = train_sample["fg"]    # PIL Image
alpha_map = train_sample["alpha"] # PIL Image (grayscale)

# Test split
test_sample = ds["test"][0]
fg_image   = test_sample["fg"]
alpha_map  = test_sample["alpha"]
trimap_map = test_sample["trimap"]

License

Non-commercial research use only.

1. Available for non-commercial research purposes only.
2. Images obtained from the Internet; authors not responsible for content.
3. No reproduction, duplicate, copy, sell, trade, or resell for commercial purposes.
4. No further public distribution except internal use within a single organization.
5. Authors reserve the right to terminate access at any time.

Citation

@inproceedings{cai2022TransMatting,
  title={TransMatting: Enhancing Transparent Objects Matting with Transformers},
  author={Cai, Huanqia and Xue, Fanglei, and Xu, Lele and Guo, Lili},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2022},
}