Robust mouse tracking in complex environments using neural networks

Both the training data and trained models used in the paper are found here.

Dataset description

Information for each dataset falls into 3 folders. Filenames portray the dataset split used in the paper that they belong to (eg Training_1.png or Validation_1.png).

• Ref/*.png: Input image (image before annotation)
• Seg/*.png: Segmentation image. Values = 0 are background. Values > 0 are mouse.
• Ell/*.txt: Ellipse-fit data. Data is tab-delimited as follows:
  • X Center of Ellipse (px)
  • Y Center of Ellipse (px)
  • Minor Axis Length of Ellipse (px)
  • Major Axis Length of Ellipse (px)
  • Angle Direction (Degrees). 0 is down with + values going counter-clockwise.

Trained Model Description

We also release models trained on all the subsets of training data we share. Each trained model was trained using our code over on Github: https://github.com/KumarLabJax/MouseTracking

Brief descriptions of the training subsets

Please read the associated paper for additional detail. A brief summary of the environment is added here:

Standard Open Field Strain Survey

We annotated 16234 training and 568 validation images of a single mouse in the same open field. The mouse can be one of multiple coat colors, but visually appears as a black, light-grey, or white color. In the case the mouse’s posture created a poor ellipse-fit, portions of the mouse were removed (such as tail) to enable a good ellipse-fit.

24Hr Open Field Dataset

We annotated 2099 training and 93 validation images of a single mouse in the same open field listed above augmented with bedding and a food container. All mice in this experiment appear black on video. There are 2 states, with visible light and with only infrared. The infrared-only imaging contains much higher visual noise.

KOMP Open Field Dataset

We annotated 1000 training and 83 validation images of a single mouse in JAX’s KOMP2 open field arena. All mice have a black coat color.

Test Ground Truth Dataset

To test the robustness of our system against conventional trackers that build a background model from multiple frames in a video, we re-sampled video a 20 minute video at 1 frame per second and annotated all the resulting frames (1179-1200 frames). We did this for the 6 environments in the paper of varying difficulty (Black, Gray, Piebald, Albino, 24Hr, KOMP2). The format of this data follows a DataSubset_FrameNumber format instead of Training/Validation_FrameNumber format.