Update README.md
Browse files
README.md
CHANGED
@@ -15,4 +15,74 @@ tags:
|
|
15 |
pretty_name: free-align-concept_covered_6M
|
16 |
size_categories:
|
17 |
- 1M<n<10M
|
18 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
pretty_name: free-align-concept_covered_6M
|
16 |
size_categories:
|
17 |
- 1M<n<10M
|
18 |
+
---
|
19 |
+
|
20 |
+
|
21 |
+
|
22 |
+
|
23 |
+
# 📦 Freeze-Align Dataset
|
24 |
+
|
25 |
+
The **Freeze-Align Dataset** (`concept_coverage_laion_6m`) is a curated collection of high-quality image-text pairs designed to facilitate efficient multimodal alignment using frozen unimodal encoders. This dataset supports the research presented in our CVPR 2025 paper, **"Harnessing Frozen Unimodal Encoders for Flexible Multimodal Alignment"**, enabling models to achieve CLIP-level performance with significantly reduced computational resources.
|
26 |
+
|
27 |
+
The dataset is curated from LAION-400M through a concept-balanced selection of captions, leveraging caption-to-image-prototype similarity to ensure diverse and semantically rich image-text pairs. The code and resources for curating this dataset are available in our [GitHub repository](https://github.com/mayug/freeze-align), enabling further research into concept coverage and reducing computational requirements for modality alignment.
|
28 |
+
|
29 |
+
## 📄 Paper
|
30 |
+
|
31 |
+
**Title:** Harnessing Frozen Unimodal Encoders for Flexible Multimodal Alignment
|
32 |
+
**Authors:** Mayug Maniparambil, Raiymbek Akshulakov, Yasser Abdelaziz Dahou Djilali, Sanath Narayan, Ankit Singh, Noel E. O'Connor
|
33 |
+
**Conference:** CVPR 2025
|
34 |
+
**Paper:** [arXiv:2409.19425](https://arxiv.org/abs/2409.19425)
|
35 |
+
**Code:** [GitHub Repository](https://github.com/mayug/freeze-align)
|
36 |
+
|
37 |
+
## 📊 Dataset Statistics
|
38 |
+
|
39 |
+
- **Total Samples:** 6,000,000 image-text pairs
|
40 |
+
- **Source:** Curated from LAION-400M using concept-balanced selection via caption-to-image-prototype similarity.
|
41 |
+
- **Image Resolution:** Variable; standardized during preprocessing
|
42 |
+
- **Text Language:** Primarily English
|
43 |
+
- **Data Format:** Parquet files with fields: `image_url`, `caption`, `embedding_vector`, `similarity_score`
|
44 |
+
- **License:** CC-BY 4.0
|
45 |
+
|
46 |
+
## 🧪 Usage
|
47 |
+
|
48 |
+
This dataset is intended for training and evaluating multimodal models that align visual and textual representations. It is particularly useful for research in:
|
49 |
+
|
50 |
+
- Multimodal representation learning
|
51 |
+
- Cross-modal retrieval
|
52 |
+
- Zero-shot image classification
|
53 |
+
- Efficient training with frozen encoders
|
54 |
+
- Representational similarity studies
|
55 |
+
|
56 |
+
To load the dataset using the Hugging Face `datasets` library:
|
57 |
+
|
58 |
+
```python
|
59 |
+
from datasets import load_dataset
|
60 |
+
|
61 |
+
dataset = load_dataset("mayug/concept_coverage_laion_6m")
|
62 |
+
```
|
63 |
+
|
64 |
+
## 📂 Dataset Structure
|
65 |
+
|
66 |
+
Each entry in the dataset includes:
|
67 |
+
- `image_url`: URL to the image
|
68 |
+
- `caption`: Associated textual description
|
69 |
+
- `similarity`: Cosine similarity score between image and text embeddings
|
70 |
+
- `IMGNET_CLASS`: One of 2754 ImageNet-derived classes the datapoint is assigned to
|
71 |
+
- `SCORE`: Cosine similarity score indicating the datapoint's association with the assigned IMGNET_CLASS
|
72 |
+
|
73 |
+
## 📬 Citation
|
74 |
+
|
75 |
+
If you use this dataset in your research, please cite our paper:
|
76 |
+
|
77 |
+
```bibtex
|
78 |
+
@inproceedings{maniparambil2025harnessing,
|
79 |
+
title={Harnessing Frozen Unimodal Encoders for Flexible Multimodal Alignment},
|
80 |
+
author={Maniparambil, Mayug and Akshulakov, Raiymbek and Djilali, Yasser Abdelaziz Dahou and Narayan, Sanath and Singh, Ankit and O'Connor, Noel E},
|
81 |
+
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
|
82 |
+
year={2025}
|
83 |
+
}
|
84 |
+
```
|
85 |
+
|
86 |
+
---
|
87 |
+
|
88 |
+
For more details and updates, please visit our [GitHub Repository](https://github.com/mayug/freeze-align).
|