Abstract
This paper introduces a robust, majority-voting-based classification scheme to filter unwanted content—specifically pornography, violence, and "ballot stuffing"—on video-sharing social networks. The authors demonstrate that using "bags of visual features" based on spatiotemporal descriptors (motion-aware features) significantly outperforms traditional static image features, achieving high accuracy across three highly diverse and challenging datasets.
Introduction
Video-sharing social networks, such as YouTube or Dailymotion, are frequently targeted by non-collaborative user behaviors. These behaviors include the upload of inappropriate content (pornography and violence) and "ballot stuffing"—the practice of posting repetitive, low-effort video replies to artificially inflate a video's popularity. Automated filtering of such content is challenging due to the "semantic gap" between low-level pixel data and high-level human concepts, as well as the extreme diversity of user-generated content. Existing methods often rely on static features, such as skin color detection, which frequently produce false positives in benign contexts like beach scenes or sports. This research proposes a feature-agnostic filtering pipeline designed to handle the complexities of real-world social media video.
Methodology
The proposed system utilizes a simple yet effective ensemble method based on majority voting. The process involves several key stages:
Element Extraction and Feature Representation
Videos are first decomposed into constituent elements, including shots, frames, and keyframes. The system then employs a Bag of Visual Features (BoVF) approach. In this framework, local features are extracted and quantized into a "visual vocabulary."
The authors evaluate various feature types:
- Static Local Features: SIFT, HueSIFT, and PCA-SIFT.
- Global Features: Color and Hue histograms.
- Spatiotemporal Features: Space-Time Interest Points (STIP), which capture motion-aware information.
Classification via Majority Voting
A separate linear Support Vector Machine (SVM) is trained for each feature type. During the inference phase, each SVM evaluates the video elements and casts a binary vote (e.g., "Inappropriate" vs. "Appropriate"). The final classification for the video is determined by a simple majority of these votes. This ensemble approach effectively smooths out noise from individual misclassified frames.
Experimental Setup
Due to the lack of standard public datasets for these specific social network problems, three custom datasets were curated:
- Pornography (800 videos): Includes 400 pornographic and 400 non-pornographic videos. The non-pornographic set contains "difficult" examples (e.g., wrestling, beach scenes) specifically chosen to challenge skin-detection algorithms.
- Ballot Stuffing (1,000 videos): Comprises 500 legitimate replies and 500 spam/stuffing replies manually annotated from YouTube.
- Violence (216 videos): Features 108 violent and 108 non-violent videos from diverse sources, including CCTV and shaky handheld footage.
Results and Discussion
The experimental results highlight the superiority of motion-aware features over static ones for high-level semantic classification.
- Pornography Detection: The STIP-based model achieved a 91.3% True Positive Rate (TPR) with an 8.7% False Positive Rate (FPR), outperforming commercial baselines like PornSeer Pro, which struggled with the "difficult" non-pornographic samples.
- Ballot Stuffing: Using STIP alone yielded a 98.0% TPR and a 12.6% FPR. While the TPR is high, the FPR suggests that fully automated deletion might be too aggressive; instead, the system could be used to de-weight spam in recommendation algorithms.
- Violence Detection: The STIP-BoVF model achieved a perfect 100% accuracy, significantly outperforming static SIFT features (80.9% TPR). Motion descriptors proved highly effective at distinguishing violent actions from visually cluttered but non-violent scenes.
Conclusion
This study demonstrates that a majority-voting ensemble of linear SVMs, particularly when utilizing spatiotemporal features, provides a robust solution for filtering unwanted content in video-sharing networks. Motion patterns are a stronger indicator of semantic content than static visual features or color information. Future research will focus on multimodal approaches, integrating audio, text, and social interaction data to further enhance classification performance and address broader categories of spam.
Summary updated