AIFacon Detection

AIFacon Detection is a product direction in planning/development. This page describes proposed scope and trials, not an available search system. No release date, price, hardware compatibility or performance is confirmed.

Usage Context and Data Preparation

AIFacon Detection aims to help find information produced by object detection through an index linked to original images or clips. Users could describe what they seek instead of opening every video interval. This is a new product direction, not confirmation that an available delivered version already ingests, indexes or searches footage.

The proposed pipeline separates ingestion, detection, event aggregation, indexing and retrieval. An index should identify events, cameras, times, objects, observable attributes and processing versions. Adding embeddings for semantic retrieval needs trials with actual data; a vector must not be treated as a complete representation of every detail or relationship in an image.

The user example is “find a woman in a red shirt carrying a box at approximately 13:00.” The person description comes from the searcher, not proof of gender identity or personal identity in the footage. Distinguish observable appearance, uncertainty and required human review. Search must not silently become automatic identification or unsupported conclusions about a person.

Approximately 13:00 does not specify a date, timezone or acceptable interval around that time. Ask the user to confirm these before bounded retrieval. The red shirt and carrying relation must belong to the same person: separate detections of a person, red clothing and a box do not satisfy the request, and proximity does not establish carrying.

The proposed search uses structured fields for explicit constraints such as cameras, dates, times, object classes or measured values, with semantic retrieval ranking eligible candidates. NLP interpretation should produce a validated query structure, not directly executed model-generated SQL. Similarity scores must not bypass access boundaries, time constraints or other required conditions.

Results should link events to original evidence with actual observation times. Selected images or clip intervals could support human verification of relationships, but descriptions must not be invented when evidence is insufficient. Distinguish no matches, incomplete indexing, missing sources and denied access without revealing information that the user is not entitled to know.

Authorization needs to cover candidate retrieval, ranking, previews and clip playback. Indexes and embeddings are not unrestricted data stores. Design deletion, permission changes and data versions consistently with original evidence, and treat captions or text in images as untrusted data rather than instructions that can change system behavior or access rules.

Before confirming capabilities, trial reference questions and events across multiple cameras and days, separating development from held-out evaluation. Measure event-level relevance, retrieval coverage, query time, index freshness and human review effort. Design and acceptance criteria remain project decisions, not published product benchmarks, service-level commitments or a guaranteed date of availability.