Research map / Vision-language and multimodal / Human and embodied modeling
LUWA Dataset: Learning Lithic Use-Wear Analysis on Microscopic Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the LUWA dataset of 23,130 microscopic images (texture and heightmap, 20x and 50x) for lithic use-wear analysis and benchmarks CNN, ViT, DINOv2, ConvNeXt, few-shot ProtoNet and GPT-4V on worked-material classification.
Problem
Identifying the worked material (wood, bone, antler, plant, etc.) from microscopic wear traces on stone tools is a hard, underexplored vision-for-science classification problem with confounders from wear mechanism and microscopic imaging, and few public datasets exist.
Approach
Builds an open-source microscopic lithic use-wear dataset spanning machine and human wear processes, two magnifications and two sensing modalities (texture and heightmap), and uses it to study how pre-trained backbones, linear probing, magnification, modality and few-shot ProtoNet generalize compared with human experts.
Key results
- Best fully supervised top-1 accuracy of 67.05% from linear probing on ImageNet-1K pre-trained ResNet152 with heightmap, 24 partitions, and 20x+50x magnification.
- Human archaeologists scored 43.75% in a 9-way-5-shot test, while several pre-trained models exceed 59.5% under some configurations.
- DINOv2 shows the most stable performance across granularity, magnification and sensing modality; DINO-base reaches 69.67% in 9-way-5-shot 50x texture style few-shot setting (Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use texture images captured at 50X rather than other magnification and modality configurations | pre-trained models with few-shot ProtoNet classification · LUWA few-shot classification | accuracy | Helped |
| increase support examples from 5-shot to 20-shot | DINO-base · LUWA 9-way few-shot, 20X texture | accuracy: 61.39 -> 69.67 | Helped |
| use 50X texture instead of 20X texture | DINO-base · LUWA 9-way-5-shot | accuracy: 61.39 -> 33.07 | Hurt |
| use 50X heightmap instead of 20X heightmap | DINO-base · LUWA 9-way-5-shot | accuracy: 42.39 -> 47.52 | Helped |
| mix 20X and 50X images instead of using 20X texture alone | DINO-base · LUWA 9-way-5-shot texture | accuracy: 61.39 -> 28.23 | Hurt |
| use texture instead of heightmap | few-shot classification models · LUWA few-shot classification | accuracy | Mixed |
| use linear probing rather than fine-tuning or training from scratch | fully supervised image classification models · LUWA | top-1 accuracy | Helped |
| increase image granularity from fewer partitions to 24 partitions | fully supervised image classification models · LUWA | top-1 accuracy | Helped |
| use higher magnification | fully supervised image classification models · LUWA | top-1 accuracy | Helped |
| mix data from different magnifications | fully supervised image classification models · LUWA | top-1 accuracy | Helped |
| use heightmaps rather than texture | fully supervised image classification models · LUWA | top-1 accuracy | Mixed |
| use GPT-4V for few-shot image classification | GPT-4V · LUWA 9-way-5-shot | accuracy | Hurt |
Surprising findings
- Linear probing outperforms full fine-tuning and training from scratch, which the authors attribute to catastrophic forgetting and dataset size.
- GPT-4V performs poorly on few-shot classification (e.g., 37.78% in 9-way-5-shot on 20x texture is below some pre-trained models) despite following expert reasoning.
Limitations
- Dataset lacks unworn regions and so cannot support segmentation or detection yet.
- Only a limited set of natural materials and man-made categories is covered.
- GPT-4V few-shot results are based on a small case study.
Experimental setup
Datasets: LUWA
Benchmarks: LUWA
Metrics: top-1 accuracy
Baselines: SIFT+FVs ResNet50 ResNet152 ConvNeXt-tiny ConvNeXt-large ViT-H GPT-4V Human experts
Builds on: DINOv2 ProtoNet Grad-CAM
Closest papers by meaning
- OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? (ICLR 2025 · similarity 0.84)
- Archaeoscape: Bringing Aerial Laser Scanning Archaeology to the Deep Learning Era (NeurIPS 2024 · similarity 0.84)
- Infusing Synthetic Data with Real-World Patterns for Zero-Shot Material State Segmentation (NeurIPS 2024 · similarity 0.82)
- Towards Automated Petrography (NeurIPS 2025 · similarity 0.81)
- 3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping (AAAI 2025 · similarity 0.81)
- Fine-Grained Perception in Panoramic Scenes: A Novel Task, Dataset, and Method for Object Importance Ranking (AAAI 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.