Computer vision, defined

The terms between a folder of images and a deployed model, in plain language: annotation shapes, dataset formats, training, metrics, and inference.

Start free Read the docs

Annotation & labeling

The shapes and labels you draw on an image to teach a model what to look for.

Annotation: An annotation is a label attached to an image that records what is present and, for most types, where it is - a box, polygon, keypoint, or class. Annotations are the supervised signal a computer vision model learns from. Auto-annotate with SAM3
Bounding box: A bounding box is an axis-aligned rectangle - x, y, width, height - that marks where an object sits in an image. It is the standard annotation for object detection. Detection datasets and models
Oriented bounding box (OBB): An oriented (rotated) bounding box adds an angle to a box so it can tightly wrap objects that are not axis-aligned, such as aerial imagery, ships, or rotated text.
Polygon: A polygon is a closed multi-vertex outline that traces an object’s exact shape. Polygons are the ground-truth input for instance segmentation and are far tighter than a box for irregular objects.
Polyline: A polyline is an open, multi-point line used to annotate linear features that have no area - lane markings, cracks, wires, or road centerlines.
Keypoint: A keypoint is a single labeled point (x, y) marking a landmark such as a joint or corner. Keypoints of one object share an instance id and connect through a class skeleton to form a pose.
Segmentation mask: A segmentation mask assigns a class to every pixel of an image (or of one object), giving pixel-precise boundaries instead of a coarse box.
Semantic segmentation: Semantic segmentation labels every pixel with a class but does not separate individual objects - all the "car" pixels are one region, whether there is one car or ten.
Instance segmentation: Instance segmentation labels every pixel with both a class and an object instance, so two overlapping cars are two distinct masks. It combines detection and segmentation.
Class: A class is a category a model can predict - "car", "pedestrian", "defect". The set of classes is the vocabulary you define for a dataset and the model learns.
Auto-annotation: Auto-annotation is model-assisted labeling: a foundation model such as SAM3 proposes masks or boxes that a human accepts or corrects, turning hours of manual work into minutes of review. How auto-annotation works
Active learning: Active learning is a labeling strategy that surfaces the most informative unlabeled images first - the ones the current model is least sure about - so you reach a target accuracy with fewer labels.
Ground truth: Ground truth is the set of human-verified, correct annotations for an image. It is what a model is trained against and what its predictions are scored against during evaluation.

Datasets & formats

How labeled images are organized, split, and exchanged between tools.

Dataset: A dataset is a collection of images together with their annotations, used to train and evaluate a model. Its quality and balance usually matter more than the choice of model architecture. Browse public datasets
COCO format: COCO is a widely used JSON annotation schema that captures bounding boxes, segmentation polygons, and keypoints in one file. Most detection and segmentation tooling reads or writes it.
YOLO format: The YOLO format stores one plain-text file per image, each line a class id plus a normalized center-x, center-y, width, and height. It is the native input for YOLO-family training.
Pascal VOC: Pascal VOC is an older per-image XML annotation format that stores object classes and bounding boxes. It is still a common interchange format for detection datasets.
Dataset split: A dataset split partitions images into training, validation, and test sets. The model learns from train, is tuned against validation, and is scored once on the held-out test set to estimate real-world accuracy.
Class imbalance: Class imbalance is when some classes have far more labeled examples than others, biasing a model toward the common classes and hurting recall on the rare ones.
Data augmentation: Data augmentation synthesizes new training examples by transforming existing ones - flips, crops, rotations, color shifts - to expand a small dataset and make the model more robust.
Export: An export is a point-in-time, format-specific package of a dataset’s images and annotations (COCO, YOLO, VOC, and more). Exports are reproducible snapshots you can archive or feed to training.

Models & training

The architectures you train and the knobs that control how they learn.

Object detection: Object detection locates and classifies each object in an image with a bounding box and a class label. It answers "what is here and where" rather than just "what is here".
YOLO / YOLOX: YOLO ("You Only Look Once") is a family of fast, single-stage object detectors. YOLOX is an anchor-free variant that pairs strong accuracy with real-time speed, making it a common production default. YOLOX vs RF-DETR
RF-DETR: RF-DETR is a transformer-based (DETR-style) detector and segmenter that predicts objects as a set, with no anchor boxes or non-maximum suppression. It trades some speed for strong accuracy on dense scenes. When to pick RF-DETR
Transfer learning: Transfer learning starts from a model already pretrained on a large general dataset and adapts it to your task, so you need far fewer labeled images to reach good accuracy.
Fine-tuning: Fine-tuning is continuing to train a pretrained model on your own dataset, adjusting its weights to your specific classes and imagery. It is the practical way to train a custom model with limited data. Train a custom model
Epoch: An epoch is one full pass of the training algorithm over every image in the training set. Training usually runs for many epochs, and too many can lead to overfitting.
Batch size: Batch size is how many images the model processes before it updates its weights once. Larger batches give smoother updates but use more GPU memory.
Overfitting: Overfitting is when a model memorizes quirks of the training set instead of learning general patterns, so it scores well on train but poorly on new images. A validation set is how you detect it.
Confidence threshold: A confidence threshold is the minimum prediction score at which a detection is kept. Raising it favors precision (fewer false positives); lowering it favors recall (fewer misses).
Non-maximum suppression (NMS): Non-maximum suppression removes duplicate, overlapping detections of the same object, keeping only the highest-scoring box per cluster. It is a standard post-processing step for detectors.

Accuracy metrics

The numbers that tell you whether a model is actually good.

IoU (Intersection over Union): IoU measures how well a predicted box or mask overlaps the ground truth: the area of their intersection divided by the area of their union. An IoU of 1.0 is a perfect match; a detection usually counts as correct above 0.5.
Precision: Precision is, of all the predictions a model made, the fraction that were correct. High precision means few false positives.
Recall: Recall is, of all the true objects present, the fraction the model actually found. High recall means few misses.
mAP (mean Average Precision): mAP is the standard object-detection accuracy metric. It averages the precision-recall curve across every class and, often, across a range of IoU thresholds - a single number summarizing overall detection quality.
F1 score: The F1 score is the harmonic mean of precision and recall, giving a single balanced number that is only high when both are high.

Inference & deployment

Turning a trained model into predictions your application can call.

Inference: Inference is running a trained model on new, unlabeled images to produce predictions - boxes, masks, or classes. It is the "use the model" step, as opposed to training it.
ONNX: ONNX (Open Neural Network Exchange) is an open, portable model format. Exporting a model to ONNX lets it run across many runtimes and hardware targets without the original training framework.
Deployment: A deployment is a hosted, always-on inference endpoint you call over an API to get predictions on demand, without managing GPUs yourself. Model deployments
Edge inference: Edge inference runs a model locally - on a device, camera, or on-prem server - instead of in the cloud, for low latency, offline operation, or data-privacy reasons. Run models locally

From glossary to working model

Pictograph is where you use them: draft labels with SAM3 and review them, train YOLOX or RF-DETR on your own classes, measure mAP and IoU, and deploy an inference API, from the app or a CLI your agent drives. Free to start, no card.

Keep learning

YOLOX vs RF-DETR Which detector architecture to pick for your accuracy and latency budget. Auto-annotation Label a dataset in a fraction of the time with SAM3-assisted annotation. Explore datasets Browse public annotated datasets and trained models to import and build on. Compare platforms How Pictograph stacks up against Roboflow, CVAT, Labelbox, V7, and more.

Build your first vision tool

$5 of free compute every month. No card required.

Start free See pricing