Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -11,7 +11,49 @@ As of this writing, ours is the only public dataset which is both fully annotate
|
|
11 |
|
12 |
It is designed to be used for controlled experiments with vision-language models.
|
13 |
|
14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
|
16 |
If you find our dataset useful, please cite our paper --
|
17 |
|
|
|
11 |
|
12 |
It is designed to be used for controlled experiments with vision-language models.
|
13 |
|
14 |
+
## What is in JANuS?[](#what-is-in-JANuS)
|
15 |
+
|
16 |
+
JANuS provides metadata and image links for four new training datasets; all of these datasets are designed for evaluation on a subset of 100 classes chosen from ImageNet-1000.
|
17 |
+
|
18 |
+
Each dataset in JANuS is either a subset or a superset of an existing dataset, and each is fully captioned and fully labeled, either using annotated or synthetic labels.
|
19 |
+
|
20 |
+
For additional details on our methodology for gathering JANuS, as well as explanations of terms like "subset matching", please refer to our paper.
|
21 |
+
|
22 |
+
1. **ImageNet-100:** A superset of ImageNet with over 50,000 newly
|
23 |
+
annotated samples, including flickr-captions and blip-captions.
|
24 |
+
|
25 |
+
2. **OpenImages-100:** A subset of OpenImages with new mappings from
|
26 |
+
OpenImages to ImageNet classes, restored original flickr-captions,
|
27 |
+
and new BLIP-captions.
|
28 |
+
|
29 |
+
3. **LAION-100:** A subset of LAION-15m with samples selected via
|
30 |
+
subset matching.
|
31 |
+
|
32 |
+
4. **YFCC-100:** A subset of YFCC-15m with samples selected via subset
|
33 |
+
matching.
|
34 |
+
|
35 |
+
## Training on JANuS[](#training-on-JANuS)
|
36 |
+
|
37 |
+
JANuS is designed to allow researchers to easily compare the effects of different labeling strategies on model performance. As such, every subset of JANuS includes at least two labeling sources.
|
38 |
+
|
39 |
+
* **idx** labels are integers, mapping to [ImageNet-1k class labels](https://deeplearning.cms.waikato.ac.nz/user-guide/class-maps/IMAGENET/)
|
40 |
+
* **caption** labels are natural language captions (usually in English), and are suitable for training VL-loss models like [CLIP](https://openai.com/blog/clip/)
|
41 |
+
|
42 |
+
For YFCC-100 and LAION-100, the idx labels are synthetic, and are generated via a simple subset matching strategy. For ImageNet-100 and OpenImages-100, the idx labels are annotated by humans.
|
43 |
+
|
44 |
+
YFCC-100, ImageNet-100 and OpenImages-100 contain captions sourced from Flickr. LAION-100 contains captions sourced from alt-text descriptions.
|
45 |
+
|
46 |
+
Additional labeling sources are available for some of the datasets in JANuS; please reference our paper for a reference key for all of the columns in the spreadsheets.
|
47 |
+
|
48 |
+
[VL Hub](https://github.com/penfever/vlhub/), a framework for vision language model training, can be used to reproduce the experiments in our paper.
|
49 |
+
|
50 |
+
## Evaluation on JANuS[](#evaluation-on-JANuS)
|
51 |
+
|
52 |
+
Evaluation methods for JANuS models are the same as those for ImageNet models, except that we evaluate only on a subset of all ImageNet classes.
|
53 |
+
|
54 |
+
For details on which classes are included in JANuS, please see metadata/in100_classes.txt in this repo.
|
55 |
+
|
56 |
+
## Citations
|
57 |
|
58 |
If you find our dataset useful, please cite our paper --
|
59 |
|