STPoseNet

Traditional methods for understanding animal behavior through videos required experimenters to rewatch videos and manually score timing and location of events. This approach is time consuming, and prone to inconsistencies between different scorers. Pose estimation methods have become widely used and can address the concerns of traditional methods by detecting underlying patterns in animal behavior. However, manual scoring can still be necessary at times, particularly in instances where video quality may be poor, such as in water maze experiments.

To address issues related to accuracy, processing speed, and annotation cost, Songyan Lv and colleagues collaborating from Guangxi University and Chongqing University have developed spatiotemporal PoseNet (STPoseNet). STPoseNet is a mouse pose estimation tool that is based on the YOLOv8 model, an established image recognition model that is capable of real-time accuracy, even in settings with complex backgrounds.  STPoseNet has high accuracy thanks to its tracking-cropping module (TCM), which uses timing information for key point detection. It also has a Kalman filter-based module, which can predict key points in consecutive frames to account for any missing data during the experiment.

STPoseNet has high accuracy, fast processing, and minimal labeling (<100 images needed for a single experimental scenario). It has been validated for mouse pose estimation both in open field and water maze settings, making it suitable for a wide array of behavioral experiments.