Datasets:
Tasks:
Visual Question Answering
Formats:
imagefolder
Languages:
English
Size:
1K - 10K
ArXiv:
License:
Update dataset card and task previews
Browse files- README.md +33 -5
- SHA256SUMS +1 -1
README.md
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
---
|
| 2 |
-
pretty_name: MultiView-Bench
|
| 3 |
task_categories:
|
| 4 |
- visual-question-answering
|
| 5 |
language:
|
|
@@ -16,12 +16,15 @@ configs:
|
|
| 16 |
drop_labels: true
|
| 17 |
---
|
| 18 |
|
| 19 |
-
# MultiView-Bench
|
| 20 |
|
| 21 |
-
|
| 22 |
-
|
| 23 |
([arXiv:2607.08970](https://arxiv.org/abs/2607.08970)).
|
| 24 |
|
|
|
|
|
|
|
|
|
|
| 25 |
## Contents
|
| 26 |
|
| 27 |
- 2,400 evaluation samples across 24 benchmark variants
|
|
@@ -35,6 +38,31 @@ It uses 3DCoMPaT++ source assets and is not part of this release. This package
|
|
| 35 |
also contains no GLTF, BIN, model asset, source code, credentials, or model
|
| 36 |
outputs.
|
| 37 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
## Load
|
| 39 |
|
| 40 |
After upload:
|
|
@@ -42,7 +70,7 @@ After upload:
|
|
| 42 |
```python
|
| 43 |
from datasets import load_dataset
|
| 44 |
|
| 45 |
-
dataset = load_dataset("suijinru/MultiViewBench
|
| 46 |
sample = dataset[0]
|
| 47 |
print(sample["benchmark"], sample["question"], sample["answer"])
|
| 48 |
print(sample["images"])
|
|
|
|
| 1 |
---
|
| 2 |
+
pretty_name: MultiView-Bench
|
| 3 |
task_categories:
|
| 4 |
- visual-question-answering
|
| 5 |
language:
|
|
|
|
| 16 |
drop_labels: true
|
| 17 |
---
|
| 18 |
|
| 19 |
+
# MultiView-Bench
|
| 20 |
|
| 21 |
+
Evaluation data for **MultiView-Bench: A Diagnostic Benchmark for
|
| 22 |
+
World-Centric Multi-View Integration in VLMs**
|
| 23 |
([arXiv:2607.08970](https://arxiv.org/abs/2607.08970)).
|
| 24 |
|
| 25 |
+
This repository currently contains the synthetic image-and-text portion of the
|
| 26 |
+
benchmark. The real-world subset is not included; see the scope note below.
|
| 27 |
+
|
| 28 |
## Contents
|
| 29 |
|
| 30 |
- 2,400 evaluation samples across 24 benchmark variants
|
|
|
|
| 38 |
also contains no GLTF, BIN, model asset, source code, credentials, or model
|
| 39 |
outputs.
|
| 40 |
|
| 41 |
+
## Task previews
|
| 42 |
+
|
| 43 |
+
Each sample asks for a relative 3D position and must be solved by integrating
|
| 44 |
+
the evidence across its views.
|
| 45 |
+
|
| 46 |
+
### Colored-object three-view task
|
| 47 |
+
|
| 48 |
+
**Question:** What is the relative position of the olive cone to the indigo
|
| 49 |
+
cube?
|
| 50 |
+
**Answer:** `(-X, -Y, -Z)`
|
| 51 |
+
|
| 52 |
+
| Front | Side | Top |
|
| 53 |
+
|:---:|:---:|:---:|
|
| 54 |
+
|  |  |  |
|
| 55 |
+
|
| 56 |
+
### 3D isometric task
|
| 57 |
+
|
| 58 |
+
**Question:** What is the relative position of the cylinder to the cone?
|
| 59 |
+
**Answer:** `(0X, 0Y, -Z)`
|
| 60 |
+
The full sample contains six views; three are shown here.
|
| 61 |
+
|
| 62 |
+
| View 0 | View 2 | View 4 |
|
| 63 |
+
|:---:|:---:|:---:|
|
| 64 |
+
|  |  |  |
|
| 65 |
+
|
| 66 |
## Load
|
| 67 |
|
| 68 |
After upload:
|
|
|
|
| 70 |
```python
|
| 71 |
from datasets import load_dataset
|
| 72 |
|
| 73 |
+
dataset = load_dataset("suijinru/MultiViewBench", split="test")
|
| 74 |
sample = dataset[0]
|
| 75 |
print(sample["benchmark"], sample["question"], sample["answer"])
|
| 76 |
print(sample["images"])
|
SHA256SUMS
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
2d178ef94980ddaf2629c3919a4cae207fa354adc42a23d23096a1ed6c60dc67 MANIFEST.json
|
| 2 |
-
|
| 3 |
5db318623cac9919e2496c5dba9be2acf87d589f09397960e10be24453263abb test/images/exp2D_three_views_2d/task_1/task_1_front.png
|
| 4 |
366d83cf773f388583f7c5a334e06f820fe01b3694b227f3d6fdb8d0f940945e test/images/exp2D_three_views_2d/task_1/task_1_side.png
|
| 5 |
d02da112f8bcc0abfe83845c2a092f2e8ed990dc7aa143f58bf0dec4303f081d test/images/exp2D_three_views_2d/task_1/task_1_top.png
|
|
|
|
| 1 |
2d178ef94980ddaf2629c3919a4cae207fa354adc42a23d23096a1ed6c60dc67 MANIFEST.json
|
| 2 |
+
379369389d3e27880c2a832031030470b04c7ab9d6c44cb88cdd3ef8ab1e4d96 README.md
|
| 3 |
5db318623cac9919e2496c5dba9be2acf87d589f09397960e10be24453263abb test/images/exp2D_three_views_2d/task_1/task_1_front.png
|
| 4 |
366d83cf773f388583f7c5a334e06f820fe01b3694b227f3d6fdb8d0f940945e test/images/exp2D_three_views_2d/task_1/task_1_side.png
|
| 5 |
d02da112f8bcc0abfe83845c2a092f2e8ed990dc7aa143f58bf0dec4303f081d test/images/exp2D_three_views_2d/task_1/task_1_top.png
|