Object detection glossary

The words you will meet in detection papers, docs and benchmarks, in plain English.

Anchor-free: A detector that predicts boxes directly from positions in the image, without a set of predefined template boxes.
AP (average precision): The area under the precision-recall curve for one class. See mAP explained.
Augmentation: Changing training images (flips, crops, colours) so the model learns the objects rather than the photos. See data augmentation.
Backbone: The first part of a detector, which turns pixels into features.
Bounding box: The rectangle around a detected object, given as corner coordinates.
Confidence: How sure the model is about a detection, from 0 to 1. See choosing a confidence threshold.
COCO: The standard detection dataset: 80 classes, about 118,000 training and 5,000 validation images. See the COCO dataset.
Epoch: One pass of training over the whole training set.
Export: Converting a trained model into a format another runtime can run, such as ONNX or TensorRT. See exporting a detector.
Fine-tuning: Training a pretrained model further on your own data. See scratch vs fine-tuning.
FP16: Half-precision numbers. Smaller and faster than FP32 on most GPUs, usually with no visible loss.
GFLOPs: Billions of floating-point operations for one image: a measure of how much computing a model needs.
Head: The last part of a detector, which predicts boxes, classes and scores.
Inference: Running a trained model on new images.
IoU (intersection over union): How much two boxes overlap: their shared area divided by their combined area.
Latency: The time from input to result for one image.
mAP50-95: Mean average precision averaged over IoU thresholds from 0.50 to 0.95: the main COCO accuracy number.
Mosaic: An augmentation that tiles four training images into one.
Neck: The middle part of a detector that mixes features from different scales.
NMS (Non-Maximum Suppression): A post-processing step that deletes overlapping duplicate boxes. fotonet does not need it. See NMS-free detection.
One-to-one assignment: A training rule that matches each real object to exactly one prediction, so the model learns not to produce duplicates.
ONNX: An open model format that runs in ONNX Runtime and many other engines, including the browser.
Precision: The share of predicted boxes that are correct.
Recall: The share of real objects that the model finds.
TensorRT: NVIDIA's inference engine, the fastest way to run a model on NVIDIA GPUs.
Tracking: Linking detections across video frames so each object keeps an ID. See tracking objects in video.
YOLO format: A common label format: one text file per image, one line per box with the class and the box centre and size from 0 to 1.