A part rolls past three cameras and gets imaged five times. If your SPC chart treats those five frames as five inspection events, you are not measuring your process — you are measuring your camera layout. That is the failure re-identification exists to prevent. On any inspection line with more than one camera, or any single camera running fast enough to capture a part in several consecutive frames, the same physical unit shows up repeatedly in the detection stream. The same physical defect shows up repeatedly too. Feed all of those observations into a statistical process control (SPC) chart as if each were an independent inspection, and the chart no longer describes the process. It describes an artefact of how you wired the vision system. Re-identification — re-ID — is the association layer that fixes this before the control chart ever reads a number. Its job is narrow and specific: link observations of the same part across cameras and time so that each physical unit contributes exactly one grounded record to the SPC stream. Done correctly, the defect-rate estimate that flows into your control limits reflects real per-unit outcomes. Done poorly, or skipped entirely, your control chart alarms on ghosts. What matters most about re-ID in practice? Re-identification answers a deceptively simple question: is the object I am looking at now the same physical object I saw a moment ago, or somewhere else on the line? In a computer vision pipeline it sits downstream of detection and tracking. A detector such as a YOLO variant or an image detection model tuned for industrial inspection finds parts in each frame. A tracker links detections across consecutive frames from one camera. Re-ID extends that association across the gaps trackers cannot bridge — between non-overlapping cameras, across occlusions, and after a part leaves and re-enters the field of view. Mechanically, most re-ID systems learn an embedding: a fixed-length feature vector that maps visually similar instances of the same object close together and different objects far apart. Two observations are judged to be the same part when the distance between their embeddings falls below a learned threshold, usually combined with spatial and temporal constraints that encode what the conveyor physically allows. A part cannot teleport upstream; two parts cannot occupy the same position; travel time between stations is bounded. These constraints matter as much as the appearance model, because on a line of near-identical parts, appearance alone is often too weak to separate one unit from the next. The distinction that trips people up: re-ID is not the same as tracking. Tracking maintains identity within a continuous observation. Re-ID re-establishes identity after continuity is lost. In practice they run together — a re-ID capable pipeline is really multi-object tracking on an inspection line plus a cross-gap association step — but the failure modes are different, and conflating them hides the specific integrity risk this article is about. Why re-ID matters for SPC: the double-counting problem SPC works because control limits are computed from the process’s own variation. When a defect rate drifts beyond those limits, the chart signals a genuine shift, and operations acts. The entire value of the method rests on one assumption: each point on the chart is an independent, correctly attributed measurement of one unit. Break that assumption and the mathematics quietly turns against you. Consider what happens when a single defective part is imaged by three cameras and each frame is logged as a separate inspection event. The defect now appears three times in the count. If a good part is imaged three times too, its “inspections” inflate the denominator. Neither the numerator nor the denominator is honest, and the ratio they produce is a function of imaging geometry, not process behaviour. Two concrete distortions follow. First, the defect-rate estimate is biased — often upward, because defects that are visually salient tend to be captured by more cameras and across more frames than clean surfaces. Second, and more damaging, the variance of the estimate is wrong. Duplicate observations are highly correlated, so counting them as independent shrinks the apparent variance and pulls the control limits inward. Tight, artificially narrow limits mean the chart alarms constantly. In our experience with multi-camera inspection cells, this is the most common root cause when a line team reports that their new SPC dashboard “cries wolf” from day one (observed pattern across CV inspection engagements; not a benchmarked rate). The correct frame is that re-ID makes the input stream statistically honest before SPC reads it. With clean association, one part produces one record, its defect status resolved by fusing what every camera saw into a single per-unit verdict. The control chart then reflects the process, and its limits mean what operations thinks they mean. What goes wrong when re-ID fails Re-ID failures split into two directions, and they produce opposite pathologies on the control chart. Naming them explicitly helps line engineers recognise which one they are looking at. Re-ID failure mode What happens mechanically SPC symptom What operations experiences Under-association (same part logged as several units) Duplicate observations counted as independent inspections Inflated defect rate, artificially narrow control limits Frequent false-alarm line stops; chart distrusted and eventually ignored Over-association (two distinct parts merged into one record) Real inspection events collapsed; some defects hidden inside a merged unit Understated defect rate, missed shifts Genuine process drift slips past the chart; escapes reach the customer Mis-association (defect attributed to the wrong part) Correct counts but wrong per-unit attribution Rate roughly right, traceability broken Root-cause analysis chases the wrong station or batch The false-alarm case is the one teams notice first, because a line stop is loud and expensive. It is also the one most directly tied to the ROI of getting re-ID right: every duplicate detection that survives into the count is a candidate false alarm, and removing them cuts the stoppage rate driven by imaging artefacts rather than real defects. The over-association case is quieter and more dangerous — a chart that under-counts defects looks healthy right up to the moment an escape triggers a customer complaint. A robust deployment has to guard against both, which is why we treat re-ID quality as a two-sided validation problem rather than a single accuracy number. Object re-ID versus person re-ID — and why we scope to parts Ask an internet search engine about re-identification and most of what surfaces is person re-ID: matching individuals across surveillance cameras. The techniques share a mathematical backbone — learned embeddings, distance thresholds, cross-camera matching — but the engineering problem on an inspection line is different in ways that matter, and we deliberately scope our manufacturing work to part-level association only. Person re-ID contends with enormous intra-class variation: the same person changes pose, clothing, lighting, and viewpoint dramatically, and the population is open-ended. Part re-ID has the opposite problem. Parts of the same SKU are nearly identical by design, so appearance features that separate people are almost useless for separating one widget from the next widget behind it. What saves part re-ID is exactly what person re-ID cannot rely on: a rigid, known physical geometry. The conveyor speed, station spacing, camera trigger timing, and part pitch are all measurable, so temporal and spatial constraints do most of the association work, with appearance embeddings resolving the residual ambiguity. Scoping to parts also keeps the deployment on the right side of a clear boundary. Person re-ID carries privacy, consent, and surveillance implications that have no place in a defect-inspection cell. Our concern is a manufactured unit’s identity across the line, resolved from the geometry of the line itself — not people. That scoping is not a limitation we apologise for; it is what makes the problem tractable and the system auditable. How to link observations across cameras and frames The association pipeline that produces clean per-part records has a recognisable shape, and each stage carries its own rationale. Detect and segment per frame. A detector or instance segmentation model built for manufacturing inspection localises each part and its candidate defects in every frame. Segmentation masks give re-ID cleaner appearance features than bounding boxes, because the background conveyor is excluded from the embedding. Track within each camera. A tracker maintains identity across consecutive frames from a single camera, absorbing the easy, high-confidence associations before anything harder is attempted. Model the line geometry. Encode conveyor speed, inter-station distance, and part pitch as a motion model. This turns “which part is this?” from an open guess into a bounded prediction — the system knows roughly where and when a given unit should reappear at the next station. Match across gaps with appearance plus constraints. Compute embedding distance between a new observation and predicted candidates, then accept the match only if it also satisfies the spatial and temporal constraints. Appearance breaks ties; geometry sets the feasible set. Fuse to one per-unit record. Once all observations of a unit are associated, resolve its defect status by fusing the views — a defect seen by any camera with sufficient confidence marks the part — and emit a single record to the SPC stream. That final fusion step is where the integrity payoff lands. It is the moment five frames across three cameras become one honest data point. How do you validate re-ID before trusting the control limits? You do not get to assume re-ID is working and then wonder why the chart misbehaves. Association quality is measurable, and it must be measured before the SPC layer is trusted, because every downstream limit is computed on the assumption that the input stream is clean. The practical validation is a labelled walk-through: run a known set of parts through the cell, some with seeded defects at known stations, and check that the pipeline produces one record per physical unit with the correct defect attribution. The two numbers that matter mirror the two failure directions — a duplication rate (units split into multiple records) and a merge rate (distinct units collapsed into one). A monitoring harness then watches these continuously in production, because line conditions drift: a bumped camera, a conveyor speed change, or a new part variant can all degrade association silently. This is precisely where re-ID meets the reliability layer. Association logic is one of the inspection artefacts a hardened deployment must sign against for auditable per-part outcomes, and its correctness is what feeds the hub’s time-to-detect-on-drift metric. Clean association keeps the input stream honest; an honest input stream is what lets the control chart detect real process drift quickly instead of chasing imaging artefacts. If you are wiring multi-camera inspection into SPC and want the association layer scoped as an explicit integrity step, that is a conversation our computer vision practice has repeatedly — and the services engagement model is built to challenge the scope before writing a line of matching code. FAQ How does re-ID work, and what does it mean in practice for CV inspection? Re-identification links observations of the same physical part across cameras and time using a learned embedding — a feature vector that places different views of the same unit close together — combined with spatial and temporal constraints from the line’s geometry. In practice it sits downstream of detection and tracking, re-establishing a part’s identity after continuity is lost between non-overlapping cameras or across occlusions, so each unit produces one grounded record. Why does re-identification matter for SPC — how does double-counting distort the control chart? SPC assumes each chart point is an independent measurement of one unit. When one part is imaged by several cameras and each frame is logged as a separate event, defects and inspections are counted multiple times, biasing the defect-rate estimate and — because duplicates are correlated — shrinking the apparent variance so control limits become artificially narrow. Re-ID collapses those duplicate observations into one honest record before SPC reads it. How do you link observations of the same part across multiple cameras and frames? Detect and segment parts per frame, track within each camera, model the conveyor geometry (speed, station spacing, part pitch) as a motion model, then match across gaps using embedding distance filtered by spatial and temporal constraints. Finally, fuse all observations of a unit into a single per-part record whose defect status combines what every camera saw. What is the difference between object re-ID for tracking parts versus person re-ID, and why scope to part-level only? Person re-ID handles huge intra-class variation across an open population and carries privacy and surveillance implications. Part re-ID faces the opposite problem — parts of one SKU are nearly identical — but benefits from rigid, measurable line geometry that does most of the association work. We scope to part-level association only because it is what makes the problem tractable and auditable, and because a defect-inspection cell has no business doing person identification. How does clean re-ID association feed the hub’s time-to-detect-on-drift metric? Clean association keeps the SPC input stream statistically honest — one part, one record, correct defect attribution — so control limits reflect the process rather than the camera layout. An honest input stream lets the control chart detect a genuine process shift quickly instead of being drowned in false alarms, which is exactly what the time-to-detect-on-drift metric measures. What goes wrong in an SPC chart when re-ID fails — false-alarm line stops or inflated defect rates? Both, in different directions. Under-association counts one part as several units, inflating the defect rate and narrowing limits until the chart triggers frequent false-alarm line stops. Over-association merges distinct parts, understating the rate and hiding genuine drift until an escape reaches the customer. Mis-association keeps counts roughly right but breaks per-part traceability, sending root-cause analysis to the wrong station. How do you validate that re-ID is associating correctly before trusting the control limits? Run a labelled walk-through — known parts, some with seeded defects at known stations — and confirm the pipeline produces one record per physical unit with correct attribution, measuring a duplication rate and a merge rate that mirror the two failure directions. A monitoring harness then tracks those numbers continuously in production, because a bumped camera, a speed change, or a new part variant can degrade association silently. Where this leaves the line engineer Re-ID is not a glamorous part of a vision system, and it rarely shows up in the demo. But it is the layer that decides whether your control chart is measuring a process or measuring a wiring diagram. Before you argue about control-limit formulas or alarm thresholds, ask the prior question: does each point on the chart correspond to exactly one physical part? If you cannot answer that with a measured duplication and merge rate, the honest move is to treat the association layer as the failure class it is — and validate it before the SPC math is allowed to mean anything.
