Sample Yoda Benchmark Videos

Pruning Strategies

  • Temporal Pruning drops frames to reduce inter-frame redundancies using at least two strategies.
    • Uniform frame selection
    • Triggered-base frame selection
  • Spatial Pruning: reencodes video to reduce redundancies among pixels.
    • Image quality downsizing
    • Region cropping
  • Model Pruning: leverages the fact that videos often have specific object classes/scenes (e.g., traffic videos contain mostly vehicles/pedestrians with static background), and trims the full DNN to reduce compute cost while still achieving high accuracy.
    • Model selection
    • Model specialization
  • Object Level Features

  • Object Speed: the reciprocal of IoU between the bounding boxes of the same object detected in two consecutive frames, to measure how fast it moves.
  • Percentage of frames with objects: percentage of frames containing objects.
  • Per-object area: the bounding box size of each object divided by the frame size.
  • Total area of objects: fraction of pixels covered by all object bounding boxes in a frame.
  • Object arrival rate: the number of new arrival objects per second.
  • Confidence score: the confidence score of each detected object given by the full DNN.
  • Object count: the number of objects per frame.