README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- visual-question-answering
language:
- en
tags:
- Agriculture
- E-commerce
- Manufacture
- Medical
pretty_name: VisualFastMappingBenchmark
size_categories:
- 1K<n<10K
---



# VisualFastMappingBenchmark


![image/png](https://cdn-uploads.huggingface.co/production/uploads/67c7bcfdfbf67e415602cff7/W3DrX9pcPN3x5M9XI_O61.png)

## Abstract
Visual Fast Mapping (VFM) refers to the human ability to rapidly form new visual concepts from minimal examples based on experience and knowledge, a keystone of inductive capacity extensively studied in cognitive science. In the realm of computer vision, early endeavors tried to replicate this capability through one-shot learning methods yet achieving limited generalization. Despite the recent advancements in Visual Language Models (VLMs), this human-like capability still has not been acquired. We introduce a novel benchmark, designed to evaluate the VFM ability in realistic industrial scenarios. Our paper and accompanying code will be publicly available online soon.

## Benchmark Construction
![image/png](https://cdn-uploads.huggingface.co/production/uploads/67c7bcfdfbf67e415602cff7/TdtNu4nDAj2kUUhWVKWxP.png)

In the previous years, plenty of high-quality datasets for perception or classification tasks on various domain have been established. Our benchmark mainly focuses on four significant industries, including agriculture, manufacturing, medicine and e-commence, whose tasks require understanding of professional vertical fields. More than 30 thousands concept images from 31 datasets have been collected as the raw data.
we employed a three-stage pipeline to curate candidate query images from raw data, ensuring the benchmark's difficulty, diversity, and quality. First, a difficulty filter was applied to exclude samples deemed insufficiently challenging, using five mainstream models as judge. Next, to promote diversity, we utilized a CLIP visual encoder to extract image features, followed by k-means clustering to sample 1,050 representative images per industry. Finally, a manual review ensured the clarity and answerability of the selected queries, resulting in a high-quality, diverse, and appropriately challenging dataset. After the whole process, 4,200 images of 512 concepts by 171 tasks have been collected in VFM Bench, as demonstrated in below Table.

|   Industry   |  Dataset Num.|  Task Num.  |  Concept Num.  | Avg. Category Per Task |
|----------|----:|---------:|---------:|---------:|
|Manufacture | 8 |     1050| 36 | 110 | 3.91  |
|M-Commerce | 11 |     1050| 22 | 109 | 7.34  |
|Agriculture | 9 |     1050| 11 | 48 | 6.77 |
|Medical | 3 |     1050| 102 | 246 | 2.37  |

## Dataset Files Introduction

The test.jsonl file only shows 4020 0-shot data entries. 5-shot example datas can be found in the 
VisualFastMappingBenchmark.zip file.


## Licensing Information
The dataset is distributed under the CC-BY-NC-SA 4.0 license.