RWTD - Real World Texture Boundary Dataset

A dataset of 253 image crops (256x256), each containing exactly two adjacent textures separated by a boundary. Every sample includes per-texture binary segmentation masks and natural-language texture descriptions.

Dataset Description

Each sample contains:

Field	Type	Description
image	Image	256x256 RGB crop
boundary_mask	Image	Binary boundary between the two textures
texture_a_mask	Image	Binary mask for texture A
texture_b_mask	Image	Binary mask for texture B
texture_a	string	Natural-language description of texture A
texture_b	string	Natural-language description of texture B
original_texture_a	string	Short original label for texture A
original_texture_b	string	Short original label for texture B
crop_name	string	Sample identifier
oracle_points_a	string	JSON-encoded list of [x,y] point prompts for texture A
oracle_points_b	string	JSON-encoded list of [x,y] point prompts for texture B

Texture Descriptions

Each texture region is annotated with a rich 5-10 word description emphasizing discriminative visual features (material, pattern, structure). Examples:

• "smooth curved seashell with radiating ridged lines"
• "fine sandy substrate with granular particles"
• "rough wood grain with prominent raised texture"
• "quilted padded fabric with raised diamond stitching"

Usage

from datasets import load_dataset

ds = load_dataset("aviadcohz/RWTD")

sample = ds["train"][0]
print(sample["texture_a"], "—", sample["texture_b"])
sample["image"].show()

Statistics

• 253 samples
• 256x256 pixel crops
• 2 textures per image with complementary binary masks
• Rich texture descriptions (5-10 words each)
• Oracle point prompts for each texture region