500 high-resolution natural images paired with manually annotated alpha mattes across categories including animals, plants, and humans. This dataset supports the development of trimap-free matting techniques by providing diverse real-world foregrounds with complex fine-grained details.
Use Cases
- Train end-to-end neural networks for alpha matte prediction without the need for user-defined trimaps
- Benchmark image matting accuracy using the provided ground-truth alpha channels
- Analyze model performance across different semantic categories such as 'plants' or 'animals' using the dataset's diverse subjects
Strengths
- 500 high-resolution natural images with corresponding ground-truth alpha mattes
- Manual annotations covering complex foreground boundaries such as hair, fur, and leaves
- Categorized into three main subsets: animals, plants, and humans/objects
- Includes a dedicated test set for benchmarking trimap-free matting algorithms