Datasets:
Tasks:
Visual Question Answering
Modalities:
Image
Formats:
imagefolder
Languages:
English
Size:
< 1K
License:
Document redistribution review
Browse files
README.md
CHANGED
|
@@ -9,7 +9,7 @@ pretty_name: AutoV Attention-Map Training Data
|
|
| 9 |
|
| 10 |
# AutoV Attention-Map Training Data
|
| 11 |
|
| 12 |
-
This
|
| 13 |
attention-map images used by AutoV. It does not contain the original source
|
| 14 |
images.
|
| 15 |
|
|
@@ -45,7 +45,6 @@ sha256sum -c SHA256SUMS
|
|
| 45 |
## Licensing note
|
| 46 |
|
| 47 |
The metadata refers to COCO, Visual Genome, GQA, OCR-VQA, TextVQA, and LLaVA
|
| 48 |
-
data. Their original licenses and terms continue to apply.
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
complete.
|
|
|
|
| 9 |
|
| 10 |
# AutoV Attention-Map Training Data
|
| 11 |
|
| 12 |
+
This dataset repository contains the training metadata and derived
|
| 13 |
attention-map images used by AutoV. It does not contain the original source
|
| 14 |
images.
|
| 15 |
|
|
|
|
| 45 |
## Licensing note
|
| 46 |
|
| 47 |
The metadata refers to COCO, Visual Genome, GQA, OCR-VQA, TextVQA, and LLaVA
|
| 48 |
+
data. Their original licenses and terms continue to apply. Redistribution of
|
| 49 |
+
the derived attention-map images was reviewed against the applicable source
|
| 50 |
+
terms before publication.
|
|
|