Roboflow 100 introduces domain-specific benchmarks beyond common
objects in context to identify how models adapt to diverse
problems in healthcare, aerial imagery, video games, and more.
Object detection benchmarks have traditionally focused on a single
dataset for fine-tuning and evaluation such as Microsoft COCO and
Pascal VOC. While semantically diverse within their domain, these
benchmarks focus on optimizing a single metric for a single
dataset, which does not show the degree of generalization learned
by the model.
In this paper we introduce the Roboflow 100 object detection
benchmark consisting of 100 projects that span a wide array of
imagery domains and task targets. We derived our benchmark
selection from over 90000 public datasets, 60 million public
images that are actively being worked on in the open on Roboflow.
The assembly of the Roboflow 100 benchmark was inspired by the
research community naturally using Roboflow datasets to test model