Detecting Sports Spoiler Images on YouTube

Spoilers of sports matches reduce the enjoyment of time-shifted viewing. On YouTube, users who like sports often inadvertently know the outcomes of matches by seeing thumbnails of recommended sports videos. Therefore, this paper focused on YouTube video thumbnails and verified the possibility of detecting images that contain spoiler information on YouTube. We constructed a dataset of sports spoiler images comprising 4,531 thumbnails from baseball, soccer, and basketball. In addition, we proposed three detection methods: the Image-Recognition method using optical character recognition (OCR), emotion assessment, and posture assessment; the Vision-Direct method using the OpenAI Vision API only; and the Vision-Text method that judges using the spoiler dictionary for an image’s description by the OpenAI Vision API. We evaluated the accuracy of these methods, and our results indicated that the Vision-Text method achieved an accuracy of 85% in detecting spoiler images. Furthermore, the evaluation results indicated that the Vision-Text method might be the most effective for detecting spoiler images in baseball and soccer. In contrast, the Vision-Direct method seems to be the most effective in basketball.