Learn
Object detection,
explained.
From the first box to a model in production. Short, practical guides, written by the people who build fotonet.
The basics
- What is object detection?Object detection finds every object in an image and draws a box around it, with a class and a confidence. Here is how it works and where it is used.5 min read
- NMS-free detection: what NMS is, and why fotonet has noneNon-Maximum Suppression removes duplicate boxes after a detector runs. fotonet is trained to predict one box per object instead, so there is no NMS step and no IoU threshold to tune.6 min read
- Choosing a model sizefotonet comes in four sizes with the same API: nano, small, medium and large. How to pick one for your hardware and accuracy needs.4 min read
- Tracking objects in videoTracking gives each detected object an ID that stays the same from frame to frame. How detection and tracking fit together, and how to track with fotonet.4 min read
Measuring
- mAP50-95 explained: how detection accuracy is measuredPrecision, recall, IoU, AP and mAP50-95: what each one means, how COCO measures detectors, and how to read a model's accuracy number.7 min read
- Choosing a confidence thresholdThe confidence threshold decides which detections you keep. How to pick it, what fotonet's default of 0.25 means, and how to tune it on your own data.5 min read
- The COCO dataset: the benchmark behind the numbersCOCO is the dataset most detectors are trained and measured on. What is in it, why benchmarks use it, and what its numbers do and do not tell you.4 min read
Training
- How to train an object detector on your own dataA step-by-step guide to training a custom object detector with fotonet: collect images, label them, write a dataset YAML, train, validate and export.8 min read
- How to label images for object detectionGood labels make good detectors. Practical rules for drawing boxes, choosing classes, handling hard cases and keeping a dataset consistent.6 min read
- Data augmentation for object detectionAugmentation creates new training images from the ones you have. Which transforms help a detector, which hurt, and how fotonet uses them.5 min read
- Training from scratch vs fine-tuningShould you train a detector from random weights or start from a pretrained model? What each means, when to use which, and how to do both with fotonet.4 min read
Deploying
- Exporting a detector: ONNX, TensorRT, TorchScript and CoreMLHow to export a fotonet model for deployment, which format to pick for which device, and how fotonet checks that an export gives the same boxes as PyTorch.6 min read
- Object detection in the browserHow fotonet runs entirely in a web browser with ONNX Runtime Web, WebGPU or WASM, and why it keeps images private.5 min read