MaterialScope

A manually annotated object detection dataset for compound material image to sub-image region detection, in COCO format.
All 2,811 images are fully annotated and usable — images with no usable annotation (all boxes labeled unlabel, or degenerate zero/negative-size boxes) have been excluded from this release. The non-semantic unlabel and common categories have also been dropped, leaving 21 real panel-label categories.

---

Dataset Summary

Metric	Value
Total images	2,811
Total annotations	12,906
Categories	21
Avg. annotations per image	4.6
Max annotations per image	21
Annotation format	COCO JSON
Image format	JPG

---

Category Distribution

Category	Annotation Count
A	2,552
B	2,539
C	2,192
D	1,979
E	1,145
F	867
G	436
H	346
single	272
I	219
J	109
K	85
L	64
M	33
N	26
O	19
P	8
Q	5
R	4
S	3
T	3
Total	12,906

---

File Formats

The dataset ships in two parallel, equivalent forms — use whichever fits your pipeline:

Format	Files	Use case
Raw COCO	merged_coco.json + images/*.jpg	Standard COCO-style object detection tooling (pycocotools, ultralytics, custom loaders)
Parquet (🤗 datasets)	data/train-*.parquet	load_dataset("CMEG-IITR/MaterialScope") — one row per image, annotations packed as list columns, images decoded on the fly. Auto-generated from the COCO source; same 2,811 images and 12,906 boxes, just reshaped

Images are JPG, RGB. Bounding boxes are always [x_min, y_min, width, height] in absolute pixels (COCO convention) in both forms.

---

Folder / File Structure

CMEG-IITR/MaterialScope/
├── README.md
├── merged_coco.json          ← COCO annotation file (images + annotations + categories)
├── images/                   ← 2,811 JPG images
│   ├── alloy_img1.jpg        ← file_name prefix marks the source annotation batch
│   ├── ceramic_img1.jpg          (alloy, ceramic, ceramic_ug, ni, polymer, single, steel, thinflim)
│   └── ...
└── data/                     ← auto-generated Parquet shards for the 🤗 datasets viewer / load_dataset()
    ├── train-00000-of-00005.parquet
    ├── ...
    └── train-00004-of-00005.parquet

The images/ filename prefix (e.g. steel_, ceramic_ug_) is not a dataset field — it's carried over from the eight original per-material annotation batches that were merged into this release, kept only to guarantee unique filenames.

---

Column / Metadata Definitions

Parquet / load_dataset() schema — one row per image; annotations on that image are packed into same-length list columns (index i of each list refers to the same box):

Column	Type	Description
image	Image	Decoded JPG image
image_id	int64	Unique image ID, matches images[].id in merged_coco.json
width	int64	Image width in pixels
height	int64	Image height in pixels
bbox	list[list[float64]]	One [x_min, y_min, width, height] box per annotation, absolute pixels
category	list[string]	Human-readable category label per box (same order as bbox)
category_id	list[int64]	Numeric category id per box (1–20 = A–T, 21 = single)
area	list[float64]	Box area in px² (width × height)
iscrowd	list[int64]	COCO crowd flag per box; always 0 — this dataset has no crowd regions

Bespoke fields (dataset-specific, not part of the standard COCO/datasets spec — defined here since they aren't self-explanatory):

• category A–T: these are panel-position labels, not material or content classes. Each compound (multi-panel) figure has its sub-images labeled A, B, C, … in reading order by the source publication; the label identifies which sub-panel a box marks, not what the panel depicts.
• category = single: used when the source image is not a compound figure — it's already a single panel. The box spans the full image (bbox ≈ [0, 0, width, height]).
• image_id: unique within this dataset only; it is a re-assigned sequential ID generated when the 8 source batches were merged, and does not correspond to any external/original annotation-tool ID.

---

Annotation Format

Annotations follow the standard COCO format:

{
  "images": [
    { "id": 1, "file_name": "img1.jpg", "width": 1721, "height": 781 }
  ],
  "annotations": [
    {
      "id": 1,
      "image_id": 1,
      "category_id": 21,
      "bbox": [x, y, width, height],
      "area": 1318264.48,
      "iscrowd": 0
    }
  ],
  "categories": [
    { "id": 1, "name": "A" },
    { "id": 2, "name": "B" },
    ...
  ]
}

Bounding boxes are in [x_min, y_min, width, height] format (COCO standard).

---

Loading the Dataset

With Python (raw COCO)

import json
from PIL import Image

with open("merged_coco.json") as f:
    coco = json.load(f)

# Build a lookup: image_id → annotations
from collections import defaultdict
ann_by_image = defaultdict(list)
for ann in coco["annotations"]:
    ann_by_image[ann["image_id"]].append(ann)

# Load an image and its annotations
img_info = coco["images"][0]
image    = Image.open(f"images/{img_info['file_name']}")
anns     = ann_by_image[img_info["id"]]
print(f"{img_info['file_name']}: {len(anns)} annotations")

With pycocotools

from pycocotools.coco import COCO

coco = COCO("merged_coco.json")
img_ids = coco.getImgIds()
ann_ids = coco.getAnnIds(imgIds=img_ids[0])
anns    = coco.loadAnns(ann_ids)

With Ultralytics (YOLO training)

pip install ultralytics

from ultralytics.data.converter import convert_coco

convert_coco(
    labels_dir=".",
    save_dir="yolo_dataset",
    use_segments=False,
)

---

License

This dataset is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

• Free to use for research and non-commercial purposes
• You must give appropriate credit when using or sharing this dataset
• Derivatives must be shared under the same license
• Commercial use is not permitted

Full license text: https://creativecommons.org/licenses/by-nc/4.0/

---

Citation

If you use this dataset in your research, please cite it as:

@article{ghosh2026unlocking,
  title={Unlocking the Visual Record of Materials Science: A Large-Scale Multimodal Dataset from Scientific Literature},
  author={Ghosh, Subham and Tiwari, Shubham and Ibrahim, Mohammad and Tewari, Abhishek},
  journal={arXiv preprint arXiv:2606.29667},
  year={2026}
}