The Google Research team has just tied first with Microsoft Research in Microsoft’s COCO (Common Objects in Context) image captioning contest where entrants were required to incorporate the Microsoft COCO dataset with their own programming to produce image recognition results that appeared as close to how a human would describe a picture as possible. Each entrant’s results were graded on their similarity to the human generated results, average correctness and amount of details in the description itself.
Microsoft COCO is a new image recognition, segmentation, and captioning dataset that is designed to recognize multiple objects and sections of an image while distinguishing their unique context. The dataset can create five separate descriptions of the image which has several uses, though the most obvious one being the improvement of image search results in search engines like Google or Microsoft’s Bing. For more details on Microsoft COCO, further challenges and download information, make sure to check out its official website.