LabOS Segmentation Dataset

A curated instance segmentation dataset of laboratory equipment that foundation models typically struggle with, and used the following to cover their gaps for LabOS tasks — including vortex genies, eppendorf tubes, multi-tube racks, colored caps, and fine-grained sub-parts like rack holes, tube tops, and mixer plates.

Annotations are provided in both COCO JSON and YOLO polygon formats.

---

Why This Dataset?

General-purpose vision models fail on lab equipment for several reasons:

• Repetitive, nearly-identical sub-objects — racks with dozens of uniform holes challenge, most foundation models have failed at, both detection and counting.
• Transparent / translucent materials — eppendorf tubes and caps have subtle visual boundaries.
• Fine-grained part segmentation — distinguishing a vortex genie top plate from its body, or an orange cap top from its barrel, requires part-level understanding that VLMs lack.
• Domain specificity — lab bench imagery is severely underrepresented in web-scraped pre-training data.

---

Dataset Statistics

Split Summary

Split	Images	Annotations
Train	251	2,989
Validation	63	682
Total	314	3,671

Split ratio: ~80 / 20 (train / val).

Annotations per Category

Category	Train	Val	Total
14ml rack hole	1,263	59	1,322
4 way rack 50ml hole	575	292	867
50ml eppendorf tube	235	85	320
14ml round bottom tube top	172	7	179
50Ml eppendorf orange cap	135	44	179
50Ml eppendorf orange cap top	106	37	143
50Ml 4 way rack	85	35	120
Vortex Genie 2	83	25	108
Vortex Genie Top Plate	67	17	84
50Ml eppendorf blue cap	49	31	80
Vortex Genie Hole	60	17	77
50Ml eppendorf blue cap top	39	25	64
50Ml eppendorf cap	47	3	50
50Ml eppendorf cap top	40	3	43
14ml rack	33	2	35
Total	2,989	682	3,671

---

File Structure

dataset-2/
├── images/                    # 314 PNG images (mixed 1280x720 and 1920x1200)
├── labels/                    # polygon segmentation (.txt, one per image)
├── annotations.json           # COCO format — all images
├── annotations_train.json     # COCO format — training split
├── annotations_val.json       # COCO format — validation split
├── dataset.yaml               # dataset config
└── demo_imgs/                 # Annotated visualization examples

---

Annotation Format

COCO JSON — bounding boxes + polygon segmentation masks per instance.

YOLO TXT — one file per image, each line:

<class_id> x1 y1 x2 y2 ... xN yN

Coordinates are normalized to [0, 1]. Annotations were created and exported from CVAT.

---

Example Visualizations

Color	Category
Red	Vortex Genie 2
Orange	Vortex Genie Hole
Yellow	Vortex Genie Top Plate
Green	50ml eppendorf tube
Blue	50Ml eppendorf cap
Purple	50Ml rack
Teal	rack 50ml hole
Gold	14ml round bottom tube top
Light blue	14ml rack hole
Violet	14ml rack
Dark orange	Orange cap / orange cap top
Dark blue	Blue cap / blue cap top

Small scene — vortex hole present, 7 classes (8 instances)

Full lab scene — vortex genie + 14ml rack + 50ml tubes (113 instances)

Vortex genie + 14ml rack with holes and tube tops (44 instances)

50ml rack — blue and orange caps, rack holes, no vortex (16 instances)

Vortex top plate + orange caps + rack holes (36 instances)

Dense 50ml rack — blue, orange & generic caps with rack holes (81 instances)

Vortex genie + orange caps, no rack holes (27 instances)

Blue caps focus — rack holes and tube bodies (42 instances)

14ml rack + vortex genie — large annotation count (130 instances)

---

Pre-trained Weights

robot-segmentation.pt — YOLO segmentation model trained on this dataset, mirrored from LabOS-Robot-Runtime. Load with:

from ultralytics import YOLO
model = YOLO("robot-segmentation.pt")
results = model("images/1.png")

---

License

This dataset and accompanying weights are released for non-commercial research use under the Creative Commons Attribution-NonCommercial 4.0 license (CC-BY-NC-4.0).