FlyingFrog/MatSim_One_Shot_Material_Recogntion_Dataset

Material Recognition and Retrieval from images dataset

The goal of the dataset is zero-shot class agnostic material recognition and retrieval from images, for any material under any conditions. Meaning the ability to recognize materials and their state from a single example in any domain.

This is fine-grained recognition of the same material and NOT classification of the material to specific set of classes. It include recognizing same material, or same states of a material, base on similarity (aka which material or material states are more similar for example in case where material gradually transform like rotting apple).

This dataset is designed to address the general issue of one-shot material retrieval without restrictions on material types, settings, and environments. The main focus is on distinguishing between states of materials and identifying fine-grained categories such as rotten vs. ripe or coffee vs. cocoa.

The dataset is composed of two parts:

1. large scale synthetic train set focus on materials in general setting as well inside transparent container and materials with partial similarity and gradual transitions.
2. Real world natural images benchmarks for one-shot material recogntion focus on materials of of any type in any setting and on materials with partial similarity and gradual transitions.
3. atural image benchmark to test the ability to identify material states and subclasses using one or a few examples (Figure 1).

Paper:

One-shot recognition of any material anywhere using contrastive learning with physics-based rendering

File Structure

Benchmark_MATSIM.zip: contain the benchmark made of real-world images as described in the paper

Dataset Generation Scripts.zip: Contain the Blender 4.1 Python scripts used for generating the dataset

MatSim_object_train_split_**.zip: Contain a subset of the synthetics dataset for images of CGI images materials on random objects as described in the paper.

MatSim_Vessels_Train_**.zip : Contain a subset of the synthetics dataset for images of CGI images materials inside transparent containers as described in the paper.

See .pdf files in for details on the dataset structure. And .py files for readers.