Annotation is the unglamorous foundation of every computer vision model. This guide focuses on annotation as it applies to ecommerce product images — what types of labels you need, how to quality-control annotation work, and when it makes sense to buy a pre-annotated dataset instead of labelling from scratch.
Annotation Types for Product Images
- Image-level class label — "this image is a sofa". Cheap to produce, sufficient for classification.
- Image-level multi-label — "this image is a sofa AND it is mid-century AND it is blue". Required for attribute extraction.
- Bounding box — marks the product's location within a lifestyle image. Required for detection models.
- Segmentation mask — pixel-level product silhouette. Required for background removal and 3D reconstruction.
- Keypoints — body or product keypoints (e.g., collar, cuffs, hem for fashion). Required for pose-based attribute models.
Annotation Cost Benchmarks (2026)
| Type | Cost per image (USD) | Time per image |
|---|---|---|
| Image-level class | $0.01–$0.03 | 5–10 s |
| Multi-label attributes (10 attrs) | $0.08–$0.20 | 30–60 s |
| Bounding box | $0.05–$0.15 | 20–40 s |
| Polygon segmentation | $0.40–$2.00 | 3–15 min |
Quality Assurance Workflow
For annotation projects with more than 10,000 images:
- Run an annotator qualification test of 50 known-good images
- Set up a review queue — sample 5–10% of daily output per annotator
- Track inter-annotator agreement (IAA) — flag annotators below 85% agreement
- Run a final golden-set validation before releasing to the training pipeline
Buy vs. Build: When Pre-Annotated Data Wins
Annotation from scratch makes sense when: your taxonomy is unique, your data is proprietary, or you need tight privacy controls. In most other cases, buying a pre-annotated dataset and fine-tuning on a small proprietary batch is faster and cheaper. ImageHub's ecommerce image dataset includes category and attribute labels ready for training — no annotation pipeline required.
Download a free annotated sample to evaluate annotation quality before committing to a full purchase, or see our FAQ for licensing details.