tolgacangoz
commited on
Commit
•
0aa362a
1
Parent(s):
2d50ff9
Update README.md
Browse files
README.md
CHANGED
@@ -62,7 +62,7 @@ The [demo notebook](./Megalith_Demo_Notebook.ipynb) shows a random sample of 100
|
|
62 |
|
63 |
Based on this random sample, I would estimate the following dataset statistics:
|
64 |
|
65 |
-
* 5-7% of images may have minor edits or
|
66 |
* 1-2% of images may be copyright-constrained (watermarks or text descriptions cast doubt on the license metadata)
|
67 |
* 1-2% of images may be non-wholesome (guns, suggestive poses, etc.)
|
68 |
* 1-2% of images may be non-photos (paintings, screenshots, etc.)
|
@@ -70,7 +70,7 @@ Based on this random sample, I would estimate the following dataset statistics:
|
|
70 |
### Is 10 million images really enough to teach a neural network about the visual world?
|
71 |
|
72 |
For the parts of the visual world that are well-represented in Megalith-10m, definitely!
|
73 |
-
Projects like [CommonCanvas](https://arxiv.org/abs/2310.16825), [Mitsua Diffusion](https://huggingface.co/Mitsua/mitsua-diffusion-one), and [
|
74 |
have shown that you can train useable generative models on similarly-sized image datasets.
|
75 |
Of course, many parts of the world aren't well-represented in Megalith-10m, so you'd need additional data to learn about those.
|
76 |
|
|
|
62 |
|
63 |
Based on this random sample, I would estimate the following dataset statistics:
|
64 |
|
65 |
+
* 5-7% of images may have minor edits or annotations (timestamps, color grading, borders, etc.)
|
66 |
* 1-2% of images may be copyright-constrained (watermarks or text descriptions cast doubt on the license metadata)
|
67 |
* 1-2% of images may be non-wholesome (guns, suggestive poses, etc.)
|
68 |
* 1-2% of images may be non-photos (paintings, screenshots, etc.)
|
|
|
70 |
### Is 10 million images really enough to teach a neural network about the visual world?
|
71 |
|
72 |
For the parts of the visual world that are well-represented in Megalith-10m, definitely!
|
73 |
+
Projects like [CommonCanvas](https://arxiv.org/abs/2310.16825), [Mitsua Diffusion](https://huggingface.co/Mitsua/mitsua-diffusion-one), and [Matryoshka Diffusion](https://arxiv.org/abs/2310.15111)
|
74 |
have shown that you can train useable generative models on similarly-sized image datasets.
|
75 |
Of course, many parts of the world aren't well-represented in Megalith-10m, so you'd need additional data to learn about those.
|
76 |
|