Unlabeled In-Domain Image Corpora for Vision Pre-Training

When in-domain pre-training beats generic ImageNet initialization

In-domain pre-training tends to win when your target images differ visually from internet photos and you have far more unlabeled images than labeled ones. In fashion, models given supervised pre-training on a large in-domain attribute dataset transferred better to fashion tasks than ImageNet-initialized models [1]; that study used labels, but it shows the value of in-domain images. The language-model literature shows the same pattern: continued pre-training on unlabeled domain text improved downstream results [6], and the comparable text-side decision is covered in sourcing domain corpora for continued pre-training.

The gap is widest for imagery that generic corpora underrepresent: grayscale line-scan images of web materials, borescope frames, close-up corrosion texture, X-ray weld radiographs, shelf photos under mixed fluorescent light. Real variation in viewpoint, illumination, weather and resolution is exactly what exposes weaknesses in general vision foundation models [3]. If your downstream task looks like a stock photo, a strong public backbone may already be enough, and you should test that before buying anything.

A practical rule: start from a strong general self-supervised backbone, then run continued self-supervised pre-training on domain images, rather than training from scratch. Large self-supervised vision models such as DINOv2 were trained on curated, rebalanced image pools rather than raw scrapes, a reminder that curation and diversity matter as much as size. The same lesson applies at domain scale: curation decides how much a purchased corpus is worth.

What to specify instead of labels

The spec for an unlabeled corpus replaces label taxonomy with coverage targets: sites, devices, conditions, time span and duplication limits. Labels are not the purchasing unit, so the questions shift to "how many distinct things does this corpus actually show?" and "will it overlap my evaluation set?"

Specify these dimensions explicitly:

• Source diversity: number of distinct facilities, stores, fields or assets, with a cap on the share any single site contributes. One plant photographed 2 million times is a narrow corpus.
• Capture diversity: camera models and lens types, fixed versus handheld versus drone, resolution bands and compression. See camera, lens and lighting diversity in image datasets.
• Temporal spread: seasons, shifts, day and night, and equipment changes over years. Seasonal coverage matters for outdoor, agricultural and infrastructure imagery.
• File fidelity: native resolution, bit depth, and whether files were recompressed. Requirements by task are in image resolution, compression and file requirements.
• Metadata retained: capture timestamp, device model, site identifier (pseudonymized) and orientation. Decide which EXIF fields to keep using EXIF metadata in image training data.
• Text that already exists: inspection notes, adjuster comments or catalog copy attached to images can serve as weak captions without paid annotation; see domain captions from work records. For the broader definition of what counts as training data, see the AI training data glossary entry.

Most operational image archives were never captured for training. Expect burst sequences, video-frame dumps, near-identical fixed-camera shots, test charts and blank frames, which is why deduplication belongs in the spec rather than in your post-delivery cleanup budget.

Deduplication and curation before you pay for volume

Deduplication should run at three levels: exact, perceptual and semantic, with thresholds agreed before delivery. Raw image counts overstate the useful size of operational archives, sometimes dramatically.

1. Exact duplicates: SHA-256 of decoded pixels (not file bytes, which change with metadata edits).
2. Perceptual near-duplicates: pHash or dHash with a Hamming-distance threshold catches resized, recompressed and lightly cropped copies.
3. Semantic near-duplicates: clustering image embeddings and pruning within tight clusters removes redundant examples that hashing misses, which SemDeDup showed can cut training data with little loss of quality [2].
4. Rebalancing: retrieve pool images nearest to a small seed set of representative target images, and downsample dense clusters (the same fixed camera, the same product) so rare conditions are not drowned out.
5. Contamination check: hash and embed your held-out downstream evaluation images and remove matches from the purchased corpus, or your benchmark will flatter the pre-training. The test design mirrors testing licensed data for overlap with public web crawls.

Ask the supplier to report counts after each stage and to deliver the hash and cluster manifest. A count quoted before deduplication is not comparable across offers.

Corpus specification template

Illustrative example: invented to show structure; it does not describe an available dataset.

Rights: pre-training scope, mixed sources and the weights after the term

A pre-training license must name pre-training as a permitted use and say what happens to model weights when the term ends. Licenses written for labeled fine-tuning often restrict use to "training a model for task X"; a backbone reused across many tasks can fall outside that wording. The rights grant you need is covered in licensing data for foundation-model pre-training, and how deletion and term clauses usually read is explained in AI data license terms.

Pre-training corpora are almost always multi-source. Even a curated public corpus built for geospatial foundation-model pre-training mixes sources under different licenses [4], and audits of public datasets found license information frequently missing or wrong on hosting sites [5]. Carry a source_id and license_id on every image in the manifest so a single contributor's withdrawal or restriction can be honored without rebuilding the corpus blind.

Operational images contain people, plates and screens more often than buyers expect. As of October 2026, Texas, for example, treats a record of face geometry as a biometric identifier and requires notice and consent before capture for a commercial purpose [7], so specify face and plate exclusion or blurring and record the method. If you ship a generative system in California, AB 2013 requires posted documentation about training data [8], which is far easier to produce when every image already carries source and license fields.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Measuring whether the corpus is worth it

Prove value with a small labeled downstream set held out from the start, not with pre-training loss. Self-supervised loss curves tell you little about whether features transfer to your task.

A workable protocol:

• Baseline: linear probe and light fine-tune of a public self-supervised backbone on your labeled downstream set.
• Candidate: the same backbone after continued pre-training on a supplier sample (for example 50,000–200,000 deduplicated images).
• Metrics: linear-probe accuracy or mAP, k-NN retrieval accuracy, and fine-tune results at 1%, 10% and 100% of downstream labels. Domain pre-training usually shows its largest benefit in the low-label regime.
• Slices: report by site, device and season, since gains concentrated in one slice may mean the sample leaked or was narrow.
• Scaling check: compare two sample sizes. If doubling the deduplicated sample yields no gain, buying 10x more will probably not help either.

How many labeled images your downstream set needs is covered in how many images you need to train a computer vision model. If the result shows labels, not volume, are your bottleneck, the trade-off is analyzed in pre-labeled vs raw image datasets.

Where unlabeled domain images come from

The richest in-domain archives sit inside operating companies: inspection and maintenance photos, quality-control line cameras, field-service job photos, insurance claim imagery, catalog and merchandising photography. These rarely appear on the open web, which is the main reason they add signal beyond what generic backbones have already seen. Category-specific considerations are covered for industrial defect images, normal-only anomaly-detection sets and product catalog photos, and the full cluster is indexed at the image data hub.

SourceX sources operational datasets, including new recordings of hands-on work, from US companies that hold the described data, on request rather than from stock; categories are not inventory and a request does not guarantee a match. Every release is approved by the supplying company, and SourceX does not source scraped web content or generic CCTV or photos. Buyers describe the data they need on the SourceX buyer page; related licensing context is on license images and inspection photos for AI training.

Before delivery, ask any supplier for per-dataset diligence covering source, rights, preparation and allowed use, and for a record of how personal details were removed or replaced. No de-identification or blurring method is perfect, so check a sample yourself.

Sourcing an in-domain image pretraining dataset

SourceX looks for US businesses that hold the domain imagery you describe, reviews ownership and consents, and delivers under a license defining the records, uses, term and delivery, through private, access-controlled workflows after an executed agreement and supplier approval. Nothing is contracted until a supplier agrees, and pricing is agreed per deal. Describe the corpus you need, including volume, diversity and pre-training scope, at sourcex.si/buyers.