Condensed Movies: Story Based Retrieval with Contextual Embeddings

Contribution:

1. CMD数据集，从MoveClips下载的30k+个包含“key scene”的视频。每个视频含有description。description中包含intent、relationship、emotion、attribute、context
   
2. movie-text retrieval 的baseline的模型。利用“Expert”模型，就是将多个多种特征级联起来（The expert features are extracted using pre-trained models for speech, motion, faces, scenes and objects.），并额外增加了一个CBM模块（allows the model to learn to increase the relative weight of a past video feature as well.）