title: Defect and Small-Object Detection: Evidence Contract description: "Defect and small-object detection produces reviewable candidates, not automatic quality truth; bind the model, capture protocol, annotation or normal-reference policy, slicing or merge mapping, thresholds, and source-disjoint evaluation before any inspection or workflow decision." category: reference tags: [defect-detection, anomaly-detection, small-objects, inspection, provenance, evaluation, tiling] aliases: ["Anomaly Detection", "Defect Detection Models", "Small Object Detection"]

Defect and Small-Object Detection: Evidence Contract¶

Detection and anomaly systems identify candidates for inspection. A box, mask, heatmap, or score is not proof that a product is defective, that a visual change was introduced, or that an image is safe to publish. Bind the capture conditions, task definition, model, and review policy before using a result in an operational decision.

Choose the task explicitly¶

EfficientAD is documented in Anomalib as a student-teacher anomaly method. Its paper evaluates a particular model family on named datasets. Neither source makes an anomaly map a universal defect label, sets a safe threshold for every material, or guarantees a fixed latency or memory budget on another runtime.

Bind capture and training evidence¶

For each sample or normal reference, retain:

• original asset/capture identity, rights, and product or scene scope;
• camera, lens, distance, focus, illumination, background, and processing conditions that affect visible texture;
• task taxonomy, inclusion/exclusion examples, annotation revision, and reviewer provenance;
• model/checkpoint, preprocessing, input resolution, runtime, and threshold policy; and
• uncertainty, rejected candidates, and known confounders.

Specular highlights, compression, dust on the capture path, focus changes, background reflections, and geometry changes can look anomalous. Keep real observations separate from synthetic defects and do not use an augmentation as proof that a real-world defect detector is validated.

Treat slicing as a coordinate contract¶

SAHI documents overlapping slices, per-slice detection, and merging back into full-image coordinates. If a high-resolution inspection uses tiles, record tile geometry, overlap, input transform, coordinate mapping, merge/suppression rule, and edge handling with the prediction. A detector's tile merge does not validate an anomaly heatmap or a pixel-editing seam.

Evaluate candidates on the original image. A downsampled preview can conceal the very small object or surface artifact the system claims to flag.

Validate for the target environment¶

Split train, tuning, and test data by original asset, product/subject, capture session, and derivative chain. Holdouts must include the material, scale, lighting, and capture variations expected at release. Report task-specific error analysis, including false candidate burden and missed critical cases, instead of transferring AUROC, mAP, VRAM, or latency values from a benchmark or another device.

Review predicted regions against the source image and the approved task definition. Keep threshold changes versioned, and revalidate when camera, lighting, preprocessing, tile plan, model, or runtime changes.

Release gate¶

Publish only reviewable candidates with model/configuration provenance and a visible uncertainty/failure state. Do not automatically reject inventory, erase retouching details, or make a safety/quality verdict solely from a detector score. If the policy requires an automatic action, define its bounded authority, counter-check, audit receipt, and rollback path before activation.

Related pages¶

• tiled inference
• low vram inference strategies
• skin retouch pipeline
• synthetic dataset pipeline