GVD: general violence detection via semantic differentiation and orthogonal supervision

Violence detection algorithms hold significant potential for security systems and online video content moderation. While current video analysis methods demonstrate effectiveness in identifying anomalous behaviors within controlled environments, two critical limitations persist: (1) Insufficient generalization capacity for novel application scenarios, and (2) Discrepancy between balanced benchmark datasets and real-world imbalanced data distributions, where violent incidents exhibit extreme rarity (typically <1%) compared to routine activities. To address these challenges, we first construct the Open-World Fights (OWF) dataset, featuring a realistic 1:100 violence-to-normal behavior ratio. We then propose GVD, a generalized violence detection framework combining universal recognition capabilities with long-tail distribution adaptation. The architecture incorporates two novel components: a Semantic Differentiation Module (SDM) that mitigates background interference through foreground-background decoupling, and an Orthogonal Supervision Module (OSM) that optimizes cross-modal alignment via feature orthogonality constraints. Extensive evaluations demonstrate GVD's competitive performance against contemporary baselines across multiple benchmarks, particularly in out-of-distribution and imbalanced data settings. Practical experiments suggest the framework's robustness in specific real-world deployments, such as continuous monitoring with the OWF dataset, though scalability challenges remain.