Short answer
Fill-level and closure inspection data consists of images of filled containers, often taken at line speed, labeled for under-fill, over-fill, missing, tilted or loose caps and seal faults. Beverage, food and pharma packagers may hold it. It can support detection and evaluation; it holds no robot actions, so it cannot train container handling.
What the task looks like on a line
On a packaging line, a filled container passes a camera station, sometimes several, a few milliseconds each. One station might look at the liquid level through the container wall, with backlighting. Another looks down at the closure from above for presence, seating and tamper-band condition. A third looks at the side for cap height and tilt. Each station produces an image and a pass or reject decision, and the reject triggers a diverter.
The visual problem differs from open-surface defect inspection because the containers vary in transparency, shape and fluid. A dark liquid in a clear bottle gives a crisp meniscus; a foamy or opaque product does not. Closures vary in color, thread style and sealing method. Ground truth often comes from a downstream check such as a weigh scale, a vacuum test or a manual review, rather than the image itself.
For the broader packaging view including seal defects, see packaging and seal defects. This page narrows to level and closure. The cluster overview is the inspection, defect and failure data hub, and the whole section is mapped in the humanoid training data guide.
Targets worth separating
- Under-fill and over-fill, judged against a level or a tolerance band specified by the packager.
- Missing, crooked or high caps, where seating is the target.
- Loose or broken tamper bands and induction seals, which may need a different view.
- Contaminated threads or sprayed product on the closure.
- Wrong closure or wrong label for the product, which overlaps with label and print quality inspection.
- Container defects such as cracks or chips, covered in glass and transparent part defects.
Each target has its own camera angle and lighting, so one blended dataset can hide poor performance on a specific target. Ask for labels per target rather than a single reject flag.
A public reference and its limits
The MVTec anomaly detection dataset has 5,354 high-resolution color images in 15 categories, five textures and ten objects. Its training set contains only defect-free images, its test set includes anomalies, and it provides pixel-precise ground truth. It is released under CC BY-NC-SA 4.0 (Bergmann et al., 2021). It is not a fill-level or closure dataset, and this page does not claim it covers beverage or pharma containers. Use it only as a labeling and format reference: it shows what a defect-free training split and pixel-level test annotations look like. The non-commercial license also means you should read terms before using it in a commercial program.
What this data can support
- Presence and seating detection for closures when images have clear views.
- Level estimation for transparent containers, if backlighting and level labels are consistent.
- Anomaly detection, with defect-free packages as the training class; see anomaly detection versus supervised defect data.
- Severity and disposition models, if reject reasons exist; see defect severity and disposition.
Reading the line context
Line speed and triggering shape what an image shows. A strobe-lit, trigger-synchronized capture freezes a container; a continuous capture at the same speed may blur the closure. Rejection decisions are also affected by timing: the image used for the decision is not always the one archived, and some systems keep only rejected frames. Ask which frames are archived, and whether an image corresponds to the one that drove the decision.
A second context question is the relationship between fill level and temperature or foam. Hot-filled and carbonated products behave differently from still, cold-filled ones, and the line may compensate with different tolerance bands. Labels from a band defined for one product are not interchangeable with another's, so record the band alongside each image.
Limits: what this data cannot support
It cannot teach a robot to handle containers. Picking bottles, twisting caps, capping by hand or clearing jams needs robot or teleoperation data with grasp, torque and contact information. Images from a fixed inspection station show none of that. For the object side, see containers, totes and bins.
It does not transfer across products. A model trained on one clear beverage bottle may not carry to an amber pharma vial or a metal food can. Labels and tolerances are product-specific, and the product's formulation affects appearance.
It reflects the line's camera station. Backlit, top-down and side views are different tasks. A mobile or humanoid inspector in a different setting would see different views.
Level correctness depends on temperature, foaming and settling time at the point of capture. A "pass" at the station may drift later.
Illustrative record specification
The following is an illustrative specification with synthetic values, not an existing dataset.
# Illustrative only: synthetic values, not a real or available dataset
record:
image_id: img_1048822
station: {id: stn_B, view: backlit_side, lighting: "diffuse backlight"}
container: {product_family: family_C, material: clear_plastic, size_class: medium}
labels:
fill_state: under_fill # under_fill | within_band | over_fill
fill_ground_truth: checkweigher
closure_state: seated # missing | crooked | high | seated
tamper_band: intact
disposition: rejected
reject_reason_code: R12
lot_id: lot_0716
line_id: line_2
split_key: lot_id
Quality and acceptance criteria
Thresholds are yours to set. Check that: ground truth for level comes from a measurement independent of the image; closure labels follow a written definition with example images; there are enough rejects of each type to evaluate that type separately; rejects caused by the camera station itself (false rejects) are recorded; and splits are by lot or line so near-identical consecutive frames do not straddle train and test. Inter-inspector agreement offers a method to test label consistency.
Permissions and provenance
Pharma and food images can reveal product formulations, packaging designs, lot numbers and regulated details. Some lines run contract packaging for brand owners, so the packager may not own the content shown. Images may include label text, batch codes or brand marks. Ask who owns the packaging design, whether the brand owner must consent, and whether regulated-product records carry restrictions. Blurring codes does not by itself resolve who may license the data. This is not legal advice.
Questions to ask suppliers
- Which stations and camera views produced the images?
- Where does fill-level ground truth come from, and is it independent of the image?
- Are false rejects and missed defects recorded?
- How many product families and container types are included?
- How were closure defects defined, and are there reference images?
- Is the archive pre-filtered to rejects, or does it include random passes?
- Do images show brand marks, lot codes or contract customers' packaging?
- Can data be separated by product so restricted products are excluded?
- Can a mixed sample of passes and rejects be reviewed before licensing?
Discuss your requirements
If you need fill-level or closure images, describe the container types, inspection views and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match; SourceX follows up about potential sources. Related catalog context is on manufacturing quality records.
Tell SourceX what your evaluation set needs to cover
Sources
- The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection — defect-free training split, pixel-level annotation and license, as a format reference only.