cubert-gmbh/XMR_Industrial_Foreign_Object_Detection_Lentils

Hyperspectral foreign-object detection in lentils: the full dataset

The larger counterpart of the tutorial demo at cubert-gmbh/XMR_Demo_Industrial_Foreign_Object_Detection_Lentils. Foreign-object detection in food sorting is an inspection problem: the rejected target can be a stone, a stem, a piece of packaging, a metal shard or an insect. Here the bulk product is bag-grade lentils (Emershofer Beluga and dark green marbled) and the contaminants span seven classes (stem_k, stone, alu_shard, blue_paper, white_paper, fly, rubber). The same frames serve supervised detection (the labelled frames) and unsupervised anomaly detection (the normal-only recordings), and the two baked splits evaluate both on identical held-out frames. The whitepaper at whitepaper/lentils_hsi_whitepaper.pdf holds the acquisition protocol, the method comparison (RGB AdaCLIP, fine-tuned AdaCLIP, Dinomaly with a custom channel selector) and the limitations. The setup is a laboratory proof of concept with production-relevant design elements, not a production deployment study.

Captured with a Cubert Ultris XMR camera (61 bands per pixel, 430 to 910 nm, about 8 nm spacing, 1080 x 1000 pixels). 3 acquisition days, 15 .cu3s recordings, 1,136 frames, 696 of them with pixel-level COCO annotations across 7 classes.

Summary

Total frames	1,136
Annotated frames	696 (61.3 %)
Annotated regions	1,536
Hyperspectral cubes (.cu3s recordings)	15
Spectral resolution	61 bands per pixel, 430 to 910 nm, about 8 nm spacing, 1080 x 1000 pixels
Processing mode	Reflectance (white and dark reference recorded per session)
Splits	dinomaly: train 308, val 148, test 180; adaclip: train 808, val 148, test 180
Total size on disk	about 57.0 GB
License	Apache-2.0

Per-day breakdown

Day	Capture date	Recordings	Frames	Annotated	Regions
day2	2026-03-03	6	384	188	368
day3	2026-03-10	6	492	328	648
day4	2026-03-17	3	260	180	520
Total		15	1,136	696	1,536

Classes

id	name	regions
1	stem_k	288
2	stone	516
3	alu_shard	112
4	blue_paper	80
5	white_paper	60
6	fly	420
7	rubber	60

Class id 0 (Unlabeled) is the implicit background: every COCO file lists it in categories and no annotation carries it. A recording with no annotation file holds normal product only; its frames are in universe.csv and carry no regions.

Why hyperspectral

An RGB sensor collapses incoming light into three bands; the human eye does the same. Hyperspectral video records 61 continuous bands per pixel and frame: a material fingerprint that separates dyes, fabrics, coatings, pigments, organic from mineral matter and surface chemistry. Foreign objects that match the colour of the bulk product (small stones in brown lentils, aluminium shards under warm lighting) are often near-isoluminant in visible RGB; they reveal themselves in the near infrared (different surface scattering, different moisture) or in narrow visible bands the eye cannot resolve. The example frames below show the same cube through three-channel projections built with the Cuvis.AI channel selectors (FixedWavelengthSelector at 650, 550 and 450 nm; CIRSelector at NIR 860, R 670 and G 560 nm), each channel min-max scaled to 8 bit.

Example frames

2026_03_03_11-11-01_id0000_cir_minmax_u8	2026_03_03_11-11-01_id0000_rgb_minmax_u8	2026_03_10_10-58-55_id0000_cir_annotated

2026_03_10_10-58-55_id0000_cir_minmax_u8	2026_03_10_10-58-55_id0000_rgb_annotated	2026_03_10_10-58-55_id0000_rgb_minmax_u8

2026_03_17_11-41-54_id0040_cir_annotated	2026_03_17_11-41-54_id0040_cir_minmax_u8	2026_03_17_11-41-54_id0040_rgb_annotated

2026_03_17_11-41-54_id0040_rgb_minmax_u8

Acquisition setup

• Camera: Cubert Ultris XMR (serial 254902), operated through CuvisNEXT
• Lens: 50 mm / f2.0
• Light source: four halogen lamps in four configurations per scene arrangement: l0 all four lights, l1 front and back, l2 front and two side lights, l3 front light only
• Working distance: 46.6 cm (field of view about 12.5 x 12 cm)
• Background: blue FDA-compliant conveyor-belt material, belt stationary during capture
• Measurement mode: snapshot, one capture per lighting configuration
• Integration time: 15 ms
• White reference: 55 % gray target
• Dark reference: recorded with the lens covered

The README of each day under data/<day>/ lists its hardware, lighting, settings, scene and recordings.

Repository layout

README.md
LICENSE                                 (Apache-2.0)
NOTICE.md                               third-party credits
manifest.json                           every file with size and sha256
fetch.py                                stdlib downloader that reads manifest.json (no token)
universe.csv                            one row per frame: source,index,annotation,group[,tags]
splits/                                 baked selector splits (core DataSplitConfig, file_indices)
  dinomaly.json                         unsupervised anomaly detection (Dinomaly): trains on the normal frames only (train 308, val 148, test 180)
  adaclip.json                          supervised baseline (AdaCLIP): the annotated positives are in train (train 808, val 148, test 180)
annotations_canonical/                  per-day concatenated COCO (time-ordered global ids)
  day2_global_coco.json                 384 images, 368 regions
  day3_global_coco.json                 492 images, 648 regions
  day4_global_coco.json                 260 images, 520 regions
assets/                                 example renderings
data/
  day2/
    README.md                             the day's hardware, lighting, settings and scene
    2026_03_03_11-11-01.cu3s              2 frames
    2026_03_03_11-11-01.info
    2026_03_03_11-11-01.json              COCO, image_id = read index
    2026_03_03_11-31-31.cu3s              17 frames
    2026_03_03_11-31-31.info
    2026_03_03_11-31-31.json              COCO, image_id = read index
    2026_03_03_11-38-39.cu3s              81 frames
    2026_03_03_11-38-39.info
    2026_03_03_11-38-39.json              COCO, image_id = read index
    2026_03_03_13-58-04_1.cu3s            96 frames
    2026_03_03_13-58-04_1.info
    2026_03_03_13-58-04_1.json            COCO, image_id = read index
    2026_03_03_13-58-04_2.cu3s            136 frames, 136 labelled
    2026_03_03_13-58-04_2.info
    2026_03_03_13-58-04_2.json            COCO, image_id = read index
    2026_03_03_15-25-02.cu3s              52 frames, 52 labelled
    2026_03_03_15-25-02.info
    2026_03_03_15-25-02.json              COCO, image_id = read index
  day3/
    README.md                             the day's hardware, lighting, settings and scene
    2026_03_10_10-17-20.cu3s              84 frames, 44 labelled
    2026_03_10_10-17-20.info
    2026_03_10_10-17-20.json              COCO, image_id = read index
    2026_03_10_10-58-55.cu3s              36 frames, 36 labelled
    2026_03_10_10-58-55.info
    2026_03_10_10-58-55.json              COCO, image_id = read index
    2026_03_10_11-30-45.cu3s              120 frames
    2026_03_10_11-30-45.info
    2026_03_10_11-30-45.json              COCO, image_id = read index
    2026_03_10_12-00-18.cu3s              40 frames, 40 labelled
    2026_03_10_12-00-18.info
    2026_03_10_12-00-18.json              COCO, image_id = read index
    2026_03_10_14-32-01.cu3s              92 frames, 88 labelled
    2026_03_10_14-32-01.info
    2026_03_10_14-32-01.json              COCO, image_id = read index
    2026_03_10_15-12-17.cu3s              120 frames, 120 labelled
    2026_03_10_15-12-17.info
    2026_03_10_15-12-17.json              COCO, image_id = read index
  day4/
    README.md                             the day's hardware, lighting, settings and scene
    2026_03_17_11-11-50.cu3s              80 frames, 80 labelled
    2026_03_17_11-11-50.info
    2026_03_17_11-11-50.json              COCO, image_id = read index
    2026_03_17_11-41-54.cu3s              80 frames, 40 labelled
    2026_03_17_11-41-54.info
    2026_03_17_11-41-54.json              COCO, image_id = read index
    2026_03_17_14-38-58.cu3s              100 frames, 60 labelled
    2026_03_17_14-38-58.info
    2026_03_17_14-38-58.json              COCO, image_id = read index

Per-recording COCO json

Standard COCO, one file per recording, image_id = the read index inside the .cu3s:

{
  "info": { "recording": "<stem>", "day": "<day>", "frame_count": N, "annotation_count": M },
  "categories": [ { "id": 0..7, "name": "..." } ],
  "images": [ { "id": <read index>, "file_name": "<stem>.cu3s", "width": W, "height": H, "camera_name": "..." } ],
  "annotations": [ { "id": ..., "image_id": <read index>, "category_id": 1..7, "bbox": [x, y, w, h],
                     "segmentation": [[...]], "iscrowd": 0, "area": 0.0, "mask": {"counts": [], "size": [H, W]} } ]
}

Annotations are semantic masks, not instances: objects of the same class in one frame share a polygon contour and carry no instance ids, and the mask field is empty (polygons only). Each image record carries three extra fields for traceability: global_frame_id (the key into the day's concatenated COCO under annotations_canonical/), camera_frame_num (the raw camera frame counter, as in the .info sidecar) and camera_name (Auto_000_<n>). Five recordings hold normal product only; their json files list the frames and no regions, and their frames are the training set of splits/dinomaly.json.

The integer id of an annotation in the day-level files under annotations_canonical/ is its position in time order and is reassigned whenever those files are rebuilt (they were in the October 2026 rebuild), so it is not a stable identifier across revisions. A region is identified by (image_id, category_id, bbox), which is unchanged between revisions.

universe.csv columns

column	meaning
source	relative posix path of the recording, e.g. data/day2/2026_03_03_11-11-01.cu3s
index	read position inside the .cu3s, equal to the COCO image_id
annotation	relative path of the recording's COCO json, empty for recordings without labels
group	frames sharing a value stay in one split (the four captures of one scene arrangement under the four lighting configurations form one group (for example day2_g000000); all four frames stay in one split so lighting alone cannot leak between splits)
tags	free labels of the recording, ;-separated (for example clean, fo); frames tagged with no tag (nothing is excluded) are kept out of every split

Splits

Splits ship as selector files under splits/ (core DataSplitConfig, file_indices) that resolve against universe.csv by (source, index).

file	train	val	test	for
splits/dinomaly.json	308	148	180	unsupervised anomaly detection (Dinomaly): trains on the normal frames only
splits/adaclip.json	808	148	180	supervised baseline (AdaCLIP): the annotated positives are in train

splits/adaclip.json is the supervised view: train 808 (500 annotated positives and 308 normals), val 148, test 180. splits/dinomaly.json is the unsupervised view: train 308 (the normals only), the same val 148 and test 180, with the 500 annotated positives held out of training. The two files share val and test, so a supervised and an unsupervised method are evaluated on identical frames.

How to load

List the test frames of a split:

import json
from huggingface_hub import hf_hub_download

repo = "cubert-gmbh/XMR_Industrial_Foreign_Object_Detection_Lentils"
path = hf_hub_download(repo_id=repo, repo_type="dataset", filename="splits/dinomaly.json")
sel = json.load(open(path))
test = [(s["source"], i) for s in sel["test"] for i in s["ids"]]
print(len(test), "test frames")

Read one recording and its labels:

import json, cuvis
from huggingface_hub import hf_hub_download

repo = "cubert-gmbh/XMR_Industrial_Foreign_Object_Detection_Lentils"
stem = "data/day2/2026_03_03_11-11-01"
cu3s = hf_hub_download(repo_id=repo, repo_type="dataset", filename=f"{stem}.cu3s")
labels = json.load(open(hf_hub_download(repo_id=repo, repo_type="dataset", filename=f"{stem}.json")))

cuvis.init()
session = cuvis.SessionFile(cu3s)
ctx = cuvis.ProcessingContext(session)
ctx.processing_mode = cuvis.ProcessingMode.Reflectance
m = ctx.apply(session.get_measurement(0))
cube = m.data["cube"].array            # (H, W, bands)
print(cube.shape, len(labels["images"]), "frames", len(labels["annotations"]), "regions")

Everything at once, with the registry:

uv run dataset download Industrial_FOD_Lentils

or without Cuvis.AI: python fetch.py from a checkout of this repo, or hf download cubert-gmbh/XMR_Industrial_Foreign_Object_Detection_Lentils --repo-type dataset --local-dir ./XMR_Industrial_Foreign_Object_Detection_Lentils.

Train on it with the cu3s_multi data module and a split file:

data:
  data_module: cu3s_multi
  splits:
    splits_path: splits/dinomaly.json

Citation

@techreport{raj2026lentilshsi,
  title       = {Spectral Foreign Object Detection in Lentils Using a Compact Hyperspectral Channel Selector},
  author      = {Raj, Anish},
  institution = {Cubert GmbH},
  year        = {2026},
  note        = {Whitepaper, May 2026},
  url         = {https://huggingface.co/datasets/cubert-gmbh/XMR_Industrial_Foreign_Object_Detection_Lentils/resolve/main/whitepaper/lentils_hsi_whitepaper.pdf}
}

License

Released under the Apache-2.0 license, see LICENSE. Third-party material is credited in NOTICE.md.

Contact

Recorded and processed by the AI team at Cubert GmbH: cuvis.ai@cubert-gmbh.de. Reach out for evaluation pilots or to run this methodology on your own product line.