PNV-ONOV-1K

ATTENTION: THIS DATASET DOES NOT PROVIDE ANY SOURCE OF VIDEOS.WE ONLY SHARE MODEL-EXTRACTED FEATURES AND METADATA, ENSURING THAT NO RAW OBSCENE CONTENT IS COMPRISED.The rapid growth of online video-sharing platforms has created to need advanced AI system of detecting and screening obscene or harmful visual content. Manual review of such media is slow, unsafe for moderators, and insufficient at large scale. To support the development of safe, ethical, and transparent obscenity detection models, we present PNV-ONOV-1K, a balanced dataset of 1,000 videos, consisting of 500 obscene and 500 non-obscene videos.The obscene category includes videos containing adult or sexually explicit visual patterns collected from various sources, while strict ethical supervision ensured that no personally identifiable, illegal, or harmful content involving minors was included. The non-obscene category consists of videos from everyday life\u2014sports, travel, cooking, education, entertainment, and public events\u2014some of which were intentionally selected to include high skin exposure (e.g., swimming, beach, pool scenes) to make the classification task more challenging.To ensure dataset safety and compliance with research-use guidelines, we do not provide raw videos. Instead, we extracted hidden representations (features) using state-of-the-art deep models. These features preserve the statistical patterns required for machine learning while preventing reconstruction of the original content. This makes PNV-ONOV-1K suitable for research in computer vision, content moderation, multimedia analysis, and ethical AI, while minimizing risks associated with handling explicit visual content.