Action Recognition Quiz

Answers and explanations appear as you take the quiz.

1. What fundamentally distinguishes action recognition from ordinary image classification?
   • It reasons about how pixels change across time, not just a single frame
   • It uses larger images
   • It only works on black-and-white video
   • It requires no training data
2. In a classic two-stream action recognition network, what does the second stream typically process?
   • Audio waveforms
   • Optical flow representing motion
   • Text captions
   • GPS coordinates
3. What does a 3D convolution add compared to a standard 2D convolution?
   • Color channels
   • Higher image resolution
   • A temporal (time) axis so filters span multiple frames
   • A larger batch size
4. What is the 'inflation' trick used by the I3D model?
   • Increasing video brightness
   • Adding more layers randomly
   • Compressing the video file
   • Expanding 2D image-network weights into 3D by replicating them across time
5. Why is the Something-Something dataset notable for action recognition?
   • It forces models to understand temporal direction rather than rely on scene context
   • It contains only animal videos
   • It has no labels
   • It is made of single images
6. Which technique helps action recognition models learn from unlabeled video?
   • Manual frame-by-frame labeling
   • Self-supervised masked video modeling
   • Deleting all motion data
   • Using only audio
7. How many action-template classes did the initial Something-Something dataset paper describe?
   • 17 classes
   • 1,740 classes
   • 174 classes
   • 17,400 classes
8. What pretraining strategy does VideoMAE use?
   • Training only on single labeled frames
   • Replacing video with a text transcript
   • Predicting optical flow without reconstructing masked content
   • Masking a high proportion of video tokens and reconstructing them from visible tokens