Video-LLaMA

• Paper: Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding
• GitHub Link
• Publisher: EMNLP 2023
• Author Affiliation: Alibaba Group
• Functional Division
  • Understanding
  • Generation
• Design Division
  • Tool-using
  • End-to-end
• Input Modalities $\rightarrow$ Output Modalities
  (I: Image, V: Video, A: Audio, 3D: Point Cloud, T: Text, I_D: Document understanding, I_B: Output bounding box, I_M: Output segmentation mask, I_R: Output retrieved images)
  • I+V+A+T $\rightarrow$ T
• Model Architecture
  (Input $\rightarrow$ Modality Encoder $\rightarrow$ Input Projector $\rightarrow$ LLM Backbone $\rightarrow$ Output Projector $\rightarrow$ Modality Generator $\rightarrow$ Output)
  • Modality Encoder
    • I/V: Eva-CLIP ViT-G/14
    • A: ImageBind
  • Input Projector
    • Q-Former w/ Linear Projector
  • LLM Backbone
    • Vicuna/LLaMA
  • Output Projector
    • None
  • Modality Generator
    • None
• Datasets Scale
  • Pre-training Stage
    • Not report
  • Instruction-tuning Stage
    • Not report