File size: 2,558 Bytes
ea38edb
 
 
 
8d0f584
ea38edb
8d0f584
 
 
690f58a
 
 
 
 
 
8d0f584
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
690f58a
8d0f584
 
690f58a
 
8d0f584
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
690f58a
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
---
license: other
license_name: probegen-data-terms
license_link: LICENSE
pretty_name: ProbeGen evaluation data
---

# ProbeGen evaluation data

Evaluation data for **ProbeGen**, from the paper
*Are Image Generators Zero-Shot Perceivers? A Rigorous Evaluation* (BMVC'26):
the 500 samples per benchmark used in the paper.

**Code:** https://github.com/Becomebright/ProbeGen. Use these archives with the
code repository and its frozen manifests.

| Archive | Benchmark(s) | Size | Files |
|---|---|---:|---:|
| `probegen-data-nyuv2.tar.gz` | NYUv2 depth | 275 MB | 1000 |
| `probegen-data-diode.tar.gz` | DIODE depth | 1.7 GB | 1500 |
| `probegen-data-kitti.tar.gz` | KITTI depth | 466 MB | 1000 |
| `probegen-data-refcoco.tar.gz` | RefCOCO, RefCOCO+, RefCOCOg (COCO images) | 157 MB | 976 |
| `probegen-data-ReasonSeg.tar.gz` | ReasonSeg | 913 MB | 1000 |
| `probegen-data-DRAM.tar.gz` | DRAM | 20 MB | 1000 |
| `probegen-data-pixmo-points.tar.gz` | PixMo-Points (images) | 134 MB | 500 |

**ScanNet and Manga109 are not included** because their terms forbid
redistribution. The code repository's `DATA.md` explains how to obtain them.
Referring-segmentation and Manga109 masks, queries and counts are embedded in
the code repository's `manifests/`.

## Usage

Run from the root of the [ProbeGen code repository](https://github.com/Becomebright/ProbeGen):

```bash
hf download Becomebright/ProbeGen --repo-type dataset --local-dir . \
  --include "*.tar.gz" --include SHA256SUMS
shasum -a 256 -c SHA256SUMS        # optional: verify the downloads
mkdir -p data
for archive in probegen-data-*.tar.gz; do tar -xzf "$archive" -C data; done
```

See the code repository's `DATA.md` for ScanNet and Manga109, and its
`README.md` for generation and evaluation.

## Sources and terms

The archives contain copies of files from NYU Depth V2, DIODE and KITTI (as
prepared by Marigold), MS-COCO 2014, ReasonSeg (LISA), DRAM and the images
linked by PixMo-Points. They are provided for non-commercial research evaluation
only. All rights remain with the original dataset authors and image owners, and
each file remains subject to the license or terms of use of its source. Please
cite the original datasets. Rights holders who want material removed can contact
dishangzhe@sjtu.edu.cn.

## Citation

```bibtex
@inproceedings{bmvc26_probegen,
  title={Are Image Generators Zero-Shot Perceivers? A Rigorous Evaluation},
  author={Di, Shangzhe and Wang, Zhaokai and Xie, Weidi},
  booktitle={BMVC},
  year={2026}
}
```

Please also cite the original datasets listed above.