Skip to content

DREEM io module.

The io module contains classes for storing, manipulating and generating model inputs and outputs.

There are 6 main submodules:

  1. Instance which represent detections. They are the main input into the model and contain
    • GT Track ID
    • Predicted Track ID
    • the crop of the image
    • the location as a centroid (x,y), bounding box (y_1, x_1, y_2, x_2) and (optionally) the pose (1, n, 2)
    • the features extracted by the visual encoder + spatiotemporal embedding
    • association score
  2. Frame which stores the metadata for a single frame of a video. It contains:
    • The video index
    • The video name
    • The frame index
    • The list of Instances that appear in that frame
    • The AssociationMatrix between the instances in that frame and all instances in the window
  3. Config which stores and parses the configs for training and/or
  4. AssociationMatrix which stores the GlobalTrackingTransformer output. AssociationMatrix[i,j] represents the association score between the ith query instance and the jth instance in the window.
  5. Track which stores references to all instances belonging to the same trajectory
  6. visualize which contains util functions for generating videos annotated by the predicted tracks and scores.