Raw videos, stored in the HDF5 format for the real-time DeepLabCut paper.
Each directory (except for every10
) corresponds to the ID of an animal (all are male C57BL6 adults).
train
sessions were acquired at 100 Hz without body-part estimation by DeepLabCut.
The frames from the videos were used to train deep-neural network models.
Body-part positions were estimated during test
sessions, using the model specifically
trained for the animal.
How body-part positions were evaluated for output generation was noted as the evaluation
attribute of the root entry of each HDF file.
The "every10" corresponds to the condition where acquisition was run without any animal, and trigger output was flipped after every 10 frames.
This condition was used to measure the timestamp-based inter-frame intervals being acquired with DeepLabCut-based pose estimation.
It is also possible to use it for validation of trigger output latency.