An object-detection dataset of small grains and seeds on a flat surface: corn, peanuts, popcorn corn, pumpkin seeds, sunflower seeds and wheat. Every image shows many objects of one type (from a few to several hundred per photo), and each object has a bounding box. It is released for education (a course project on object detection); see License below.

Task: object detection, 6 classes (category ids 0-5, fixed order below).
Format: COCO (bbox = [x, y, width, height] in pixels, top-left origin).
Size: 470 photos, 1,875 tiles, 27,611 tile-level boxes.

id	name
0	corn
1	peanuts
2	popcorn_corn
3	pumpkin
4	sunflower
5	wheat

Folder structure

README.md  LICENSE.md  classes.json  splits.json  stats.json  checksums.sha256
coco/
  train/  images/*.jpg  labels.json     tile-level images + COCO labels (this is what a detector trains on)
  val/    images/*.jpg  labels.json
  test/   images/*.jpg  labels.json
photos/
  val/*.jpg   test/*.jpg                full photos of the val and test splits (not tiled)
annotations/
  val_photos.json  test_photos.json     COCO labels for those full photos
scripts/
  load_example.py   verify.py           tiny helpers (see below)

Tiles (coco/*/images): photos were cut into tiles of at most 1500 x 1500 px with 20% overlap. Photos up to 1600 px on both sides are a single tile (the whole photo). Tile names are <photo_name>_x<X>_y<Y>.jpg, where X, Y is the tile's top-left corner in the photo. A box is kept in a tile only if at least 80% of its area is inside that tile and the clipped box is at least 8 px on each side. Some tiles have no boxes (kept as background examples, at most about 10% of a split).
Photos (photos/val, photos/test): the full images, so you can evaluate at photo level (run your detector on tiles, merge the boxes back using the x, y in the tile name, compare with annotations/*_photos.json). Training photos are not provided as full images, only as tiles. Full-photo sizes: 76 images of about 1.5k x 1.5k px and 55 images of 3072 x 4080 or 4080 x 3072 px.
Photo names are <class>_<NNNN>.jpg; the number is arbitrary and carries no meaning. splits.json maps each photo name to train, val or test.
All images are JPEG without EXIF/colour-profile metadata. Pixel coordinates refer to the stored pixels (no rotation flags).

Counts

Tile-level (what coco/*/labels.json contains). A grain lying in the overlap of two tiles, or cut by a tile border, can appear in more than one tile, so box counts here are not counts of distinct grains.

class	split	photos	tiles	empty tiles	boxes	mean boxes / tile	median box w x h (px)	box side min / median / max (px)
corn	train	58	102	2	2175	21.3	130 x 134	68 / 132 / 214
corn	val	8	96	4	903	9.4	139 x 129	75 / 134 / 195
corn	test	8	83	8	683	8.2	135 x 135	68 / 134 / 315
peanuts	train	55	99	3	1849	18.7	121 x 120	59 / 121 / 229
peanuts	val	11	33	3	236	7.2	124 x 130	80 / 128 / 234
peanuts	test	11	120	13	820	6.8	158 x 153	88 / 154 / 365
popcorn_corn	train	63	250	17	4288	17.2	89 x 89	40 / 89 / 162
popcorn_corn	val	11	55	2	983	17.9	84 x 86	46 / 87 / 102
popcorn_corn	test	10	21	2	605	28.8	84 x 80	39 / 84 / 116
pumpkin	train	81	411	11	4950	12.0	115 x 121	48 / 121 / 239
pumpkin	val	14	80	5	740	9.2	135 x 138	60 / 127 / 387
pumpkin	test	15	81	11	672	8.3	118 x 136	22 / 121 / 396
sunflower	train	33	165	11	2891	17.5	110 x 120	54 / 109 / 165
sunflower	val	14	25	2	559	22.4	112 x 119	54 / 107 / 138
sunflower	test	9	31	3	528	17.0	115 x 121	58 / 108 / 139
wheat	train	49	115	7	3276	28.5	72 x 74	37 / 69 / 167
wheat	val	7	73	2	579	7.9	86 x 81	60 / 82 / 150
wheat	test	13	35	0	874	25.0	77 x 74	44 / 71 / 123
all	train	339	1142		19429	17.0		37 / 105 / 239
all	val	65	362		4000	11.0		46 / 107 / 387
all	test	66	371		4182	11.3		22 / 113 / 396

Photo-level (annotations/val_photos.json, annotations/test_photos.json; one box per object, no duplicates):

class	val photos	val boxes	test photos	test boxes	mean boxes / photo (val+test)
corn	8	334	8	297	39.4
peanuts	11	191	11	344	24.3
popcorn_corn	11	506	10	480	47.0
pumpkin	14	395	15	432	28.5
sunflower	14	520	9	380	39.1
wheat	7	225	13	695	46.0

How to load

With pycocotools:

from pycocotools.coco import COCO
coco = COCO("coco/train/labels.json")
print(len(coco.getImgIds()), "tiles,", len(coco.getAnnIds()), "boxes")
img = coco.loadImgs(coco.getImgIds()[0])[0]            # img["file_name"] is relative to coco/train/images/
anns = coco.loadAnns(coco.getAnnIds(imgIds=img["id"]))  # each ann: bbox=[x,y,w,h], category_id in 0..5

With torchvision: torchvision.datasets.CocoDetection("coco/train/images", "coco/train/labels.json").

Detector frameworks that expect a "COCO folder" (for example RF-DETR) need one folder per split with the images and one JSON file. Here the JSON is labels.json next to an images/ folder; if your framework wants the JSON inside the image folder (often named _annotations.coco.json), copy or symlink it there.

scripts/load_example.py prints the counts per class and saves an overlay with the boxes of one tile. scripts/verify.py checks the integrity of the release (ids, boxes inside images, no photo in two splits, counts, checksums).

How the splits were made

Photos of the same sample group (consecutive shots of the same pile of grains, and photos with matching layouts) were kept together, and whole groups were assigned to train, val or test. Every class has several groups in each split, and no photo or group appears in two splits. The split is therefore by source photo and group, never by tile. Tiles of one photo overlap and look alike, and photos of one group look almost the same: do not re-split the tiles randomly (or re-mix train/val/test), or your validation scores will be inflated by near-duplicates. If you need a different split, group by photo name (and keep similar photos together) rather than by tile. splits.json gives the assignment per photo.

Annotation notes (please read)

The bounding boxes were generated automatically with a vision-language model (one box per object, on image tiles) and then filtered with simple size rules (boxes much smaller or much larger than the typical box of the same photo, or very elongated, were removed).
They were NOT manually verified or corrected. Expect missed grains, doubled boxes, loose or tight boxes, and boxes merging two touching grains.
Therefore any metric you compute on val/test measures agreement with these automatic labels, not accuracy against a human-verified ground truth. Report it that way.
Tile-level boxes are clipped copies of the photo-level boxes (see the tile rule above); box widths and heights are in tile pixels.

Known limitations

One class per image: no photo mixes grain types, so a model can learn "background look -> class" shortcuts. Do not expect it to separate mixed piles.
Unequal class sizes: pumpkin seeds and popcorn corn have the most photos, sunflower seeds the fewest (see the table).
Small val/test: only about 65 photos each, from a handful of independent groups per class, so per-class scores are noisy.
Limited variety: a few backgrounds, one lighting style and a limited range of viewpoints and distances. Models may not transfer to other backgrounds or lighting.
Dense piles contain touching and overlapping grains, which are hard for both the labels and the models.

Authors

Niranjan Kukkala (niranjan.reddy.wo@gmail.com)

Download

hf download niru-5/manama_course_dataset_grain_detection --repo-type dataset --local-dir grain_detection_public
python grain_detection_public/scripts/verify.py        # integrity check, prints "all checks passed"

Intended use

Teaching and coursework in object detection (training, evaluating, error analysis, and discussing label quality). Not intended for any commercial use.

License

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0): you may share and adapt the data for educational / non-commercial purposes, with attribution. See LICENSE.md.

More information

More details about the dataset (collection protocol, annotation process, benchmarks) will be released after the course.

Downloads last month: 43

Total file size:

1.62 GB

Models trained or fine-tuned on niru-5/manama_course_dataset_grain_detection

niru-5/manama_course_rfdetr_grain_detection

Object Detection • 31.8M • Updated 7 days ago • 42

Grain Detection Dataset (six classes)