Datasets:
File size: 2,913 Bytes
25acfe4 772a14b 25acfe4 772a14b b8ebf13 772a14b b8ebf13 772a14b b8ebf13 772a14b b8ebf13 772a14b | 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 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 | ---
license: mit
task_categories:
- image-classification
tags:
- shape-recognition
- benchmark
- 3d-objects
- shapenet
pretty_name: ShapeY
size_categories:
- 100K<n<1M
---
# ShapeY
ShapeY is a benchmark that tests a vision system's **shape recognition capacity**. It
consists of ~68k images of 200 3D objects rendered from
[ShapeNet](https://shapenet.org/).
ShapeY is **not a training set**. It validates that an object recognition system has
developed a genuine capacity for shape understanding, using tasks that are hard for
systems that rely on texture or contrast cues instead of shape.
Code: <https://github.com/njw0709/ShapeY>
## Contents
| Directory | Images | Description |
|---|---|---|
| `ShapeY200/<objname>/` | 68,200 total | The standard benchmark. |
| `ShapeY200CR/<objname>/` | 68,200 total | Contrast-reversed counterparts, for the CR experiment. |
200 objects across 20 categories (10 objects per category), each rendered in 31
transformation series of 11 views: 341 images per object, 68,200 per set.
**Images are grouped one directory per object** (200 directories of 341 images),
because the Hub does not allow more than 10,000 files in a single directory. The
benchmark code expects a single flat directory, so
[`scripts/download_data.py`](https://github.com/njw0709/ShapeY/blob/main/scripts/download_data.py)
flattens the tree after downloading. If you fetch the files yourself, collapse the
per-object directories before pointing `SHAPEY_IMG_DIR` at them -- filenames are
globally unique, so a flat merge is lossless.
## Filename convention
```
<category>_<shapenet_id>-<axis><NN>.png
```
* `<axis>` is a non-empty subset of the five transformation axes `x`, `y`, `p`, `r`, `w`
(translation in x and y, pitch, roll, and scale/width), always in that canonical
order — 31 combinations in total.
* `<NN>` is the view index within the series, `01`–`11`, increasing distance from the
reference view.
Example: `airplane_1021a0914a7207aff927ed529ad90a11-pr03.png` is view 3 of the combined
pitch+roll series for that airplane.
The parser for these names is `shapeymodular.utils.ImageNameHelper.parse_imgname`.
## Usage
```bash
git clone https://github.com/njw0709/ShapeY && cd ShapeY
uv sync
uv run scripts/download_data.py --variant all --out data
export SHAPEY_IMG_DIR=$PWD/data/ShapeY200/dataset
export SHAPEY_IMG_DIR_CR=$PWD/data/ShapeY200CR/dataset
```
Or directly:
```python
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="GamGyulNN/ShapeY",
repo_type="dataset",
allow_patterns=["ShapeY200/**"],
local_dir="data",
)
# -> data/ShapeY200/<objname>/<objname>-<axis><NN>.png
```
Note the per-object subdirectories; flatten them before use (see above).
## License
MIT for the benchmark itself. The renders are derived from ShapeNet — please also
observe [ShapeNet's terms of use](https://shapenet.org/terms).
|