Deeplabcut-OpenFieldArena

A Deeplabcut network trained to label mice in open field arena with topdown view.

Network configuration

The network was initialized using the pretrained network mobilenet_v2_1.0. Training dataset was augmented with the imgaug method. The network was continously trained with manually annotated videos until satisfactory results were obtained. In total, 1674 frames were labelled and ~3000000 iterations were performed.

Network performance

The network was evaluated to have a 1.13px train error and 4.82px test error, using a scale factor(train-test ratio) of 0.8. We have a short 2-min video demonstrating the accuracy of our tracking: demo

We also compared the labelling performance with a commercial video tracking software called Ethovision. As shown in speed plot below, our DLC model has comparable performance with commerical options.

Speed plots generated from DLC labelling data and Etho labelling data

Labelling configuration

These are the body parts labelled:

• Nose
• Left ear
• Right ear
• Centroid (Body center)
• Left lateral(Left hip-joint in anatomical terms)
• Right lateral (Right hip-joint in anatmoical terms)
• Tail base
• Tail end

Note: Snout, Tail base and Tail end are relative unstable key points, meaning they might be occassionally off-target in some frames.

Step-by-Step tutorial

Follow these simple steps to start running analysis on your machine with our trained network!

Step 1: Pre-processing of videos

To get the best results, please adjust the video to follow the specifications below:

• format: .avi
• speed: 1X
• resolution: 1080px x 1080px
• frame rate: 10fps

We used a free editor software called VSDC Free Video Editor:http://www.videosoftdev.com/free-video-editor/download

Note: Changing the frame rate and resolution hinders the performance of the network.

Step 2: Download the network

1. Download the folder here
2. Create a new project in DLC
3. Place the downloaded folder in the dlc-model folder in your project
4. Go to the subfolder dlc-model/iteration-0/YourprojectnameDatecreated-trainset95shuffle1.. Rename the subfolder by replacing the text Yourprojectname with your project name and Date created with the date created

• eg. OpenFieldTestJul13-trainset95shuffle1

5. You are good to go! Click the analyze video tab in the DLC GUI. If you are using our network, you should see Using snapshot-950000 for model .........../dlc-models/iteration-0/........-trainset95shuffle1 pop up in the command prompt.