نسخة أولية وصول مفتوح
MVVBench: Benchmarking 4D Reasoning in Vision-Language Models
Multi-view video understanding requires integrating spatial and temporal evidence across multiple, often non-overlapping camera streams: tracking entities as they transition between viewpoints, aligning events across time, and reasoning about latent 4D continuity rather than any single visible frame. We introduce MVVBe …