Racing Bib Number Recognition   Idan Ben-Ami              Tali Basha              Shai Avidan Abstract   We propose an automatic system for racing bib number (RBN)  recognition in natural image collections covering running races such as marathons. An RBN is typically a piece of durable paper or cardboard bearing a number as well as the event/sponsor logo. The RBN, usually pinned onto the competitor’s shirt, is used to identify the competitor among thousands of others during the race. Our system receives a set of natural images taken in running sport events and outputs the participants’ RBNs. Today, RBN identification is often done manually, a process made difficult by the sheer number of available photos. This specific application can be studied in the wider context of detecting and recognizing text in natural images of unstructured scenes. Existing methods that fall into this category fail to reliably recognize RBNs, due to the large variability in their appearance, size, and the deformations they undergo. By using the knowledge that the RBN is located on a person’s body, our dedicated system overcomes these challenges and can be applied without any adjustments to images of various running races taken by professional as well as amateur photographers. First, we use a face detector to generate hypotheses regarding the RBN location and scale. We then adapt the stroke width transform (SWT) to detect the location of the tag, which is then processed and fed to a standard optical character recognition (OCR) engine. We evaluate the contributions of each component of our system, and compare its performance to state-of-the-art text detection methods, as well as to a commercially available, state-of-the-art license plate recognition (LPR) system, on three newly collected datasets. -------------------------------------------------------------------------------------------------------------------------- Paper     [BMVC'12]    -------------------------------------------------------------------------------------------------------------------------- Code [RBNR v0.1] -------------------------------------------------------------------------------------------------------------------------- Data Contains 217 color images and ground truth RBNs per image divided into three sets, each taken from a different race. If you use these datasets in any publication, please refer to our paper. [Download]