[CVPR 2024 Highlight] LUWA Dataset: Learning Lithic Use-Wear Analysis on Microscopic Images.

Jing Zhang*, Irving Fang*, Hao Wu, Akshat Kaushik, Alice Rodriguez, Hanwen Zhao, Juexiao Zhang, Zhuo Zheng, Radu Iovita, Chen Feng

Project Website

Please visit our project website for more information, including an interactive demo with real artifacts.

Environment Setup

The project was developed on Python 3.11.5 and PyTorch 2.1.1 with CUDA 11.8.0 binaries. While you can refer to requirements.txt for more details, having PyTorch installed should be mostly enough for this project.

We utilized PyTorch 2.1.1 to access some of its exclusive features such as torch.compile() to accelerate training as much as we could. However, most (if not all) of these techniques should not affect the inference accuracy, so you should be able to perfectly replicate our results without a matching PyTorch version.

Dataset

Please visit our Hugging Face repo to access the dataset.

Please refer to transfer_learning/data_utils/data_tribology.py for how to process them.

1. we use integers to label stone that has been worked again certain material as follows:

   Material	Integer
   ANTLER	0
   BEECHWOOD	1
   BEFOREUSE	2
   BONE	3
   IVORY	4
   SPRUCEWOOD	5
   BARLEY	6
   FERN	7
   HORSETAIL	8

   Here, "BEFOREUSE" refers to a state where the stone is not polished with any material at all.

2. In the dataset,