Dataset Viewer
Auto-converted to Parquet Duplicate
dataset
string
checked_at
timestamp[s]
status
string
errors
int64
warnings
int64
settings
dict
images_checked
bool
stats
dict
checks
list
issues
list
E:\Sythetic_Data\tool-detector\synthetic_raw
2026-11-18T23:09:45
pass
0
0
{ "tiny_box": 0.01, "imbalance_ratio": 3, "duplicate_iou": 0.95 }
true
{ "images": 1500, "label_files": 1500, "boxes": 3177, "boxes_per_image_mean": 2.12, "boxes_per_image_max": 4, "empty_images": 0, "classes": [ "wrench", "screwdriver", "hammer", "pliers" ], "boxes_per_class": { "wrench": 740, "screwdriver": 999, "hammer": 728, "pliers": ...
[ { "check": "missing_label", "severity": "error", "description": "Image has no label file", "count": 0 }, { "check": "orphan_label", "severity": "error", "description": "Label file has no image", "count": 0 }, { "check": "unreadable_image", "severity": "error", "de...
[]

ToolVision Synthetic Hand Tools

Synthetic images rendered in Blender with randomized scenes, with YOLO bounding-box labels generated automatically from the 3D scene.

At a glance

Images 1,500
Boxes 3,177 (2.12 per image)
Classes 4: wrench, screwdriver, hammer, pliers
Resolution 1024x768, 768x1024
Format YOLO (class x_center y_center width height, normalized 0-1)
Validation passed (0 errors, 0 warnings)

Classes

id class boxes share
0 wrench 740 23.3%
1 screwdriver 999 31.4%
2 hammer 728 22.9%
3 pliers 710 22.3%

Files

images/              images
labels/              one .txt per image (YOLO boxes)
classes.txt          class names, line number = class id
data.yaml            Ultralytics dataset file
metadata.json        how the data was made (settings, seed, versions)
validation.json      automatic checks and statistics
quality_report.html  visual report: charts and sample images with boxes

Train with Ultralytics:

huggingface-cli download moindedroper/toolvision-synthetic --repo-type dataset --local-dir toolvision-synthetic
cd toolvision-synthetic && yolo detect train data=data.yaml model=yolo11n.pt

data.yaml uses all images for both train and val. Split off your own validation set (and test on real images) before trusting the scores.

How it was made

Rendered with ToolVision generate_tools.py v1.1 in Blender 4.5.3 LTS (Cycles), through SynthPipe.

Setting Value
Seed 42
Render size 1024x768
Cycles samples 64
HDRI environments 14
Floor textures 14
Objects per image 1 to 4
Object scale 0.85 to 1.15
Camera distance (m) 0.3 to 0.9
Camera elevation (deg, 90 = top-down) 40 to 85
Lens (mm) 26
HDRI light strength 0.2 to 1.8
Chance of a spotlight-only scene 0.35
Exposure -0.3 to 0.8
Chance of a textured floor 0.75
Chance of flipping an object 0.5
Chance of portrait framing 0.4

Label rule: full tool outline (including parts covered by other tools), clipped to image edge.

Known limitations

  • Only one 3D model each for wrench, hammer and pliers; three for screwdriver.
  • Classes are not balanced: screwdriver has 999 boxes, pliers has 710 (1.4x).
  • Objects sit mostly near the middle of the frame: 67% of box centres fall in the central 40% x 40% of the image (16% of its area). Expect weaker detection near image edges.
  • Synthetic images differ from real photos (lighting, materials, camera noise). Models trained only on this data should be tested, and ideally fine-tuned, on real images.

License

Released under cc-by-4.0. Check the license of any third-party assets (3D models, textures, HDRIs, photos) before commercial use.

Citation

@misc{toolvision_synthetic,
  title  = {ToolVision Synthetic Hand Tools},
  author = {moindedroper},
  year   = {2026},
  url    = {https://huggingface.co/datasets/moindedroper/toolvision-synthetic}
}

Card generated by SynthPipe on 2026-11-19.

Downloads last month
47