VideoJev: System One Model for Video Understanding and Interaction
VideoJev turns video tasks into scored decisions instead of decoded answers.
The paper presents a model family that scores temporal intervals and text options against video evidence, covering both temporal grounding and multiple-choice video QA. It reuses visual computation through cached inference for incoming queries and joint inference for multiple queries at once. The authors report mIoU gains over backbones on three temporal grounding benchmarks and large inference speedups, including over 100x on four video QA benchmarks. Source: HF Daily Papers' note.
The paper presents a model family that scores temporal intervals and text options against video evidence, covering both temporal grounding and multiple-choice video QA. It reuses visual computation through cached inference for incoming queries and joint inference for multiple queries at once. The authors report mIoU gains over backbones on three temporal grounding benchmarks and large inference speedups, including over 100x on four video QA benchmarks. Source: HF Daily Papers' note.
score 4