Question on Video Quality and Labeling for Networking Training

My questions concern the quality of video and labeling before training the network

Labeling:

  • Does the size of the marker used for labeling affect training or is that more for visualization purposes? Is the size of the marker important to capture the center of the object of interest as well as surrounding pixels?

Video Quality:

  • Due to the nature of our setup, our depth-of-field/focus is limited due to the lighting of our setup. How much does the background of the video and quality of the image affect tracking?

Best,

Pratik

Hello Pratik!

The size of a marker is only for visualization purposes; it is its center location that is used during training.
As for the video quality, please have a look at our paper ‘On the inference speed and video-compression robustness of DeepLabCut’, Fig.3C. The average tracking error barely increases despite a 3-order of magnitude video compression (from about 2 GB down to 1.5MB).

Cheers,
Jessy

1 Like

Perhaps additionally with regard to Video Quality, as the broad range of applications on the website show: http://www.mousemotorlab.org/deeplabcut the algorithm is strong to work in almost any circumstances.

However, the harder something is to see for you, the more labeled data might be required.