The COCO dataset: the benchmark behind the numbers
When a detector reports an accuracy, it is almost always measured on COCO. Knowing what COCO is helps you read those numbers.
What is in it
Common Objects in Context (COCO) is a public dataset of everyday photos: streets, kitchens, sports, animals. Every object from 80 classes, such as person, car, dog, chair and cup, is outlined with a box.
- train2017: about 118,000 images for training.
- val2017: 5,000 images held out for measuring.
- Objects are grouped by size: small, medium and large. Small objects are the hardest.
Why benchmarks use it
COCO is hard, varied and the same for everyone. Two detectors scored on val2017 with the official tool can be compared directly. That is why fotonet's accuracy is quoted as COCO mAP50-95 on all 5,000 validation images.
What it does not tell you
Your data is not COCO. A model that leads on COCO may not lead on X-rays, aerial photos or a factory line. Use COCO numbers to shortlist models, then measure on your own images.
Training from scratch on COCO
fotonet's models learn COCO directly from random initialization, with no pretrained backbone and no teacher model. When you train on your own data, you start from those COCO weights instead, which is much faster. See training from scratch vs fine-tuning.