Walking behavior of flies (Drosophila melanogaster)

Dataset description

The dataset consists of walking behaviors of flies, Drosophila melanogaster, that were placed in an acrylic arena that constrained them to move in a 2D plane. The flies were part of a genetic screen (not wild-type) but were examined during behavior capture and were morphologically and behaviorally indistinguishable from wild-type flies.

We placed three female flies in one arena simultaneously and filmed them from below for 3 seconds (frame rate 80Hz). This was repeated several times for different sets of flies. To extract each fly from the video separately, we tracked the centroid of each fly using Tracktor (Sridhar et al. 2019), cropped out the flies in each frame, and aligned them to point in the same direction. We then tracked 32 body parts (four joints per leg, as well as head features, thorax, abdomen, and wings), each with x- and y-directions using DeepLabCut (Mathis et al. 2018). This resulted in time series with 64 feature dimensions. We then smoothed the extracted time series using a Savitzky-Golay-Filter (Savitzky and Golay, 1964).

Dataset dimensions

number of sequences
28059

sequence length
234

keypoints (2 per keypoint)
64

frame rate: 80 Hz 

Loading the data in Python

import h5py

with h5py.File('Fly_DLC_behavior_tracking.h5', 'r') as f:
data = f['data'][:]