Bagheera Bghira
commited on
Commit
•
bae8d71
1
Parent(s):
e7a4e8b
readme
Browse files
README.md
CHANGED
@@ -1,3 +1,31 @@
|
|
1 |
---
|
2 |
license: cc
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: cc
|
3 |
---
|
4 |
+
|
5 |
+
# pseudo-camera-10k dataset
|
6 |
+
|
7 |
+
### Work in progress (incomplete dataset)
|
8 |
+
|
9 |
+
## Contents
|
10 |
+
|
11 |
+
This dataset contains 10k free images from world class photographers. The images have been resized using Lanczos antialiasing, with their smaller edge shifted to 1024px.
|
12 |
+
|
13 |
+
No images were upsampled in this dataset. The results of downsampling were manually validated to ensure limited artifacting.
|
14 |
+
|
15 |
+
The images have been captioned with CogVLM.
|
16 |
+
|
17 |
+
## Hardware
|
18 |
+
|
19 |
+
These images were captioned on a single A100-80G at a rate of roughly 3.8 seconds per image.
|
20 |
+
|
21 |
+
## Use cases
|
22 |
+
|
23 |
+
These images may be used to train models.
|
24 |
+
|
25 |
+
## Limitations
|
26 |
+
|
27 |
+
Natural photos can be highly noisy. This will be learnt by the model; that may be a desired trait.
|
28 |
+
|
29 |
+
The captions are not 100% accurate, but they are highly descriptive and close to 100%.
|
30 |
+
|
31 |
+
The images do not reflect the multitude of diversity that much larger sets would.
|