~/glossary // plain-language terms
Glossary
The terms we use across the lab, in plain words. Each one gives the business meaning first, then a precise line for engineers.
Concepts
- Cold start concept
The drop in accuracy when the system meets a new store, product, or setting with no data yet.
- Embedding concept
A numeric fingerprint of an image or text that lets a computer compare things by how similar they are.
- Logit concept
The raw score a model gives each option before it picks the highest.
- Planogram concept
The store's plan for how products should sit on each shelf.
- Realgram concept
Our method that reads what recently sat at each shelf spot to sharpen recognition.
- Vector search concept
Finding the closest items by comparing their numeric fingerprints.
Models
- AuraSR model
An AI model that upscales images to look sharper, tested here on price tags.
- CatBoost model
A model that weighs many signals together to pick the final answer.
- ConvNeXt model
A modern convolutional vision model, one of the backbones we weigh for turning photos into features.
- Flux ControlNet Upscaler model
An AI upscaler that adds detail to images, tested here on price tags.
- LLM model
A model trained on text that answers and writes in plain language.
- Swin Transformer model
A vision transformer that reads an image in shifted windows, another backbone we weigh for product recognition.
- ViT-B/32 model
A mid-size vision model we use to turn product photos into fingerprints.
- VLM model
A model that looks at an image and answers questions about it in words.
- YOLO model
The fast detector that finds each product on a shelf photo and boxes it.
Datasets
Metrics
- Cosine similarity metric
A score for how alike two product fingerprints are, higher means closer.
- IoU metric
A 0 to 1 score for how much two boxes overlap, used to match products across frames.
- On-Shelf Availability metric
The share of products actually on the shelf when a shopper looks, the number retail cares about.
- Precision metric
The share of the system's calls that turn out to be correct.
- Recall metric
The share of the things that should be found that the system actually finds.
- SSIM metric
A 0 to 1 score for how alike two images look, used to spot what changed on a shelf.
- Top-1 accuracy metric
How often the system's single best guess is the correct one.
- Variance of the Laplacian metric
A quick sharpness score for an image, higher means crisper.
Techniques
- ArcFace technique
A training method that teaches a model to pull photos of the same product together.
- DBSCAN technique
A clustering method that groups dense points and labels the loners as outliers.
- DeepSORT technique
Frame-to-frame tracking that keeps the same product labeled the same across photos.
- Far-point sampling technique
Picking a spread-out subset of points so the selection still covers the whole space.
- Fine-tuning technique
Taking a model trained on general data and training it further on your own.
- FLANN technique
A fast way to find the closest matches among many candidate points.
- KNN technique
Judging a point by the labels of its closest neighbors.
- Metric learning technique
Training that places look-alike product crops close together so new items match by similarity.
- OCR technique
Reading the text off an image, here the price and name printed on a shelf tag.
- Perceptual hashing technique
A short fingerprint of an image so near-identical pictures get near-identical codes.
- QuerySoftMax technique
The training target that teaches the ranker to pick one right product per crop.
- SHAP technique
A way to see how much each feature changed a model's decision.
- SIFT technique
A classic way to find and match distinctive points between two photos.
- Super-resolution technique
Upscaling a low-resolution image to look sharper, here tried to make small tag text readable.
- Test-time augmentation technique
Running several altered copies of one input through a model and pooling the answers for a steadier result.
Infrastructure
- Airbrush infra
An online image enhancement service, tested here on price tags.
- albumentations infra
A fast library of image augmentations for training and testing.
- Cleanlab infra
A tool that flags likely label errors in a dataset.
- CleanVision infra
A tool that finds duplicate or low-quality images in a dataset.
- FiftyOne infra
A tool for browsing and curating image datasets by hand.
- Hugging Face infra
A public hub for sharing and downloading machine-learning models.
- Numba infra
A tool that speeds up Python number-crunching by compiling it.
- Optuna infra
The tool that searches for good model settings so we do not tune them by hand.
- Qdrant infra
The searchable index where product fingerprints live, so lookups stay fast.
- torchvision infra
A PyTorch library of image models, datasets, and transforms.