What happens when your retail CV system encounters a product it has never seen? A shelf monitoring system is deployed with a trained catalogue of 1,200 SKUs. Three weeks after deployment, a seasonal promotion introduces 80 new products. The CV system has three options for handling these objects: misclassify them as the nearest known SKU, return a low-confidence “unrecognised” label, or route them to an explicit review queue. In a system without a designed unknown-object handling path, the first option is what happens by default. The model returns its best guess — the nearest class in feature space — at whatever confidence the inference produces. If that confidence is above the acceptance threshold, the misclassification enters the system’s output without any signal that it is wrong. The shelf analytics report shows planogram compliance metrics that are incorrect for the 80 new products, and no one knows until a manual audit. This is the unknown-object problem in retail CV: not the misclassification itself, which is an expected and manageable consequence of having an incomplete catalogue, but the silent treatment of misclassifications as correct decisions. Why unknown objects accumulate Retail environments have continuous catalogue churn. New products launch; existing products change packaging; regional variants enter the assortment; seasonal items rotate in and out. The rate of change varies by retailer; we have observed annual catalogue change rates in the 5–15% band across the retail CV deployments we have worked on (an observed range, not a benchmarked industry rate). For a 1,000-SKU catalogue, that is 50–150 product changes per year — roughly one per week. A model retrained on a quarterly or monthly schedule will systematically lag behind the actual assortment. During the gap between retraining cycles, the new products are in the store but not in the model. They are unknown objects. Without designed handling, they accumulate in the misclassification pool. The misclassification rate for new products is not random: new products tend to misclassify into visually similar existing products, which means misclassifications cluster within product categories. If the shelf monitoring system is used to measure category-level planogram compliance, the errors are correlated and directional — the compliance metric for the new product’s category is systematically biased for the entire period between retraining cycles. The designed unknown-object loop An unknown-object loop converts the unknown-object problem from a silent accuracy drain into a continuous improvement cycle. The loop has four stages: Stage Mechanism Output Detect Out-of-distribution (OOD) detector flags predictions where the model’s feature representation is far from all known class centres. Threshold calibrated to the known distribution. A per-prediction OOD score indicating confidence that the object is unknown. Route Predictions above the OOD threshold are not classified; they are labelled “requires review” and routed to an operator review queue with the raw image crop attached. A review queue populated with unknown-object candidates for operator labelling. Label Operators (or a structured crowdsourcing pipeline) label each unknown-object image with the correct SKU identity or confirm it as a genuinely new product requiring catalogue registration. Labelled examples for the new SKU class. Retrain When the label count for a new SKU class crosses a training threshold, the new class is added to the model’s training set and the model is updated incrementally. An updated model that now classifies the previously unknown SKU correctly. In the share-of-shelf and planogram analytics system we developed, this loop was designed as a first-class pipeline component. Products that the model had not been trained on generated OOD scores above the detection threshold and were surfaced to the operator review queue consistently, rather than misclassifying into the nearest known category. The review queue served double duty: it caught and corrected planogram compliance errors in real time, and it generated the labelled training examples that allowed the new products to be added to the model within one retraining cycle. The out-of-distribution detection implementation OOD detection for retail CV is most practically implemented at the embedding layer rather than the classification head. The classification head produces a probability distribution over known classes — it has no representation for “unknown” and its confidence values are not calibrated for OOD detection. The embedding layer produces a feature representation in a learned metric space where known classes form identifiable clusters. The most operationally straightforward OOD approach for production retail CV is nearest-neighbour distance scoring: at inference time, compute the L2 or cosine distance from the predicted embedding to the nearest known class centroid. If this distance exceeds a threshold calibrated to the training distribution, the object is flagged as potentially unknown. The threshold is set on the validation set to achieve a target true-positive OOD detection rate while keeping the false-positive rate (known objects flagged as unknown) within an operationally acceptable band; we typically calibrate to a 2–5% false-positive band, depending on the operator capacity for review (project-specific calibration, not a universal rule). This approach adds minimal inference latency (a single distance computation against pre-computed class centroids), requires no changes to the classification model architecture, and is compatible with any standard embedding-based recognition model. OOD detector training and threshold calibration: a worked procedure The nearest-neighbour distance approach is operationally simple but the calibration step is where most implementations get into trouble. The procedure below is the structure we use; the specific numbers depend on the embedding model and the catalogue. 1. Prepare the held-out class set. From the training catalogue, hold out 5–10% of classes entirely (do not train on them). These held-out classes simulate genuinely unknown objects: the model has never seen them, but their images are real product photos with the same lighting and capture conditions as the training set. A retail catalogue of 1,200 classes might hold out 60–120 classes, with 30–100 images per held-out class. 2. Train the embedding model normally on the in-distribution classes. No special OOD-aware loss is required for the nearest-neighbour distance approach. Standard contrastive or classification training on the in-distribution classes produces an embedding space where in-distribution classes cluster. 3. Compute class centroids. For each in-distribution class, compute the mean embedding of its training images. These centroids are the reference points for distance scoring at inference time. Store them as a fixed lookup table (1,000 classes × 256-dimensional embeddings = 256,000 floats, trivial memory). 4. Score the validation set and the held-out set. For each image in both sets, compute the embedding and the distance to its nearest class centroid. The validation set produces a distribution of in-distribution distances; the held-out set produces a distribution of out-of-distribution distances. Plot both distributions on the same axis. 5. Choose the threshold from the overlap region. The two distributions overlap (otherwise OOD detection would be trivial). The threshold is chosen to balance two error rates: the false-positive rate (in-distribution images flagged as OOD) and the false-negative rate (held-out images that pass through as in-distribution). A worked example: if the in-distribution distance distribution has mean 0.32 and standard deviation 0.08, and the held-out distribution has mean 0.71 and standard deviation 0.14, a threshold of 0.55 produces approximately a 2.5% false-positive rate and a 12% false-negative rate. The exact numbers depend on the embedding model; the procedure is general. 6. Validate the operator capacity assumption. The false-positive rate determines the review queue volume on in-distribution traffic. For a system processing 100,000 images per day with a 2.5% false-positive rate, that is 2,500 review items per day from in-distribution traffic alone, plus the genuinely unknown objects. If operator capacity is 1,000 review items per day, the threshold needs to be tightened (higher distance, lower false-positive rate, higher false-negative rate) and the consequences accepted. 7. Recalibrate periodically. As the catalogue grows, the in-distribution distance distribution shifts (more classes, denser embedding space, smaller gaps between centroids). Recalibrate the threshold after each major retraining cycle and after any architectural change to the embedding model. The calibration procedure should produce a numerical artefact — the chosen threshold and the validated false-positive and false-negative rates at that threshold — that becomes part of the model’s release documentation. A model deployed with an OOD detector but without a documented calibration is operating an OOD detector whose behaviour is unknown. The operational difference between a system with and without the loop A system without an unknown-object loop processes an expanding catalogue by allowing misclassification to accumulate until the next retraining cycle corrects it. The accuracy degradation is silent and directional. The team knows the system has a problem when planogram compliance metrics start diverging from manual audits. A system with a designed loop processes catalogue change continuously. The review queue volume is a direct signal of the current unknown-object rate — it tells the team exactly how many new products are in the store and not yet in the model. The loop converts catalogue dynamism from a source of accuracy degradation into a source of training data. The compound failure class that affects retail CV at scale includes unknown-object accumulation as one of its four axes. A system that does not handle unknowns explicitly is not managing the compound failure class — it is accepting it silently. The loop design converts silent acceptance into explicit management. A Production CV Readiness Assessment for retail evaluates whether the planned system has a designed unknown-object loop or is implicitly relying on retraining cadence to absorb catalogue change.
