FlyingFrog/MatSeg_Zero-Shot_Material_Segmentation_Dataset

MatSeg Dataset

The MatSeg dataset focuses on zero-shot segmentation of materials and their states identifying image regions belonging to a specific material type or state, without any prior training on that material, state, or environment.

It emphasizes complex, scattered, and sparse material boundaries, as well as soft similarity and gradual transitions between materials.

For Example:

• Regions belonging to different materials
• Regions belong to the same material but in different states:
• Wet regions on surfaces
• Scattered dust
• Minerals in rocks
• Sediment in soils
• Rotten parts of fruits
• Degraded and corrosive surface regions

---

Paper

This dataset is introduced and described in: Infusing Synthetic Data with Real-World Patterns for Zero-Shot Material State Segmentation

Training Dataset

Composed of 100,000 images, the synthetic dataset including 3D rendered scenes with realistic illumination and PBR materials as well as 2D maps. With textures boundaries extracted from real-world images and infused into synthetic data

Real-World Benchmark

The benchmark contains 1,220 annotated real-world images for evaluation and testing across a diverse set of material states and environments.

It focuses on:

• Zero-shot material segmentation (segmenting region of the image belonging to the same material without for any material type)
• Separation between different states of the same material
• Soft segmentation and partial similarity (For example regions if A and B are same material but in different states like wet/dry)
• Segmentation of materials with complex and scattered boundaries

For Example:

• Food states and spoilage
• Plant diseases and infections
• Rocks and minerals
• Soil and sediment
• Corrosion and rust
• Liquids and foam
• Worn or degraded surfaces
• Diverse unconstrained environment

File Structure

File	Description
Dataset_Documentation_And_Readers.zip	Dataset readers and technical documentation
MatSeg_Benchmark.zip	Real-world evaluation benchmark
MatSeg2D_part_*.zip	2D synthetic training set
MatSeg3D_part_*.zip	3D synthetic training set