Published in De Computis, this study by Raman Srinivasan, R. Maskeliūnas, and R. Damaševičius addresses the challenge of accurately recognizing dog poses without the use of physical markers. Pose recognition plays a central role in animal behavior research, welfare assessment, ecological monitoring, and interactive applications. Traditional methods often require large labeled datasets or intrusive markers, limiting real-world usability.
The authors propose a methodology that combines video frame extraction, deep convolutional neural network (CNN) training, and semi-supervised learning to recognize canine poses. Their approach uses a ResNet-based model that leverages both restricted labeled data and a large volume of unlabeled images, reducing the need for extensive manual annotation. This semi-supervised reinforcement framework allows the model to generalize effectively across varied environments and dog morphologies.
Sequential CNNs are used for feature localization—identifying key points of the dog’s body—and for spatio-temporal analysis, enabling the recognition of motion and posture sequences. By extracting annotations directly from image frames, the system avoids starting from scratch with a feature model, thus decreasing the computational load and dataset size requirements typically associated with pose-recognition tasks.
The methodology was validated on a dataset of more than 5,000 images depicting dogs in a wide range of poses and behaviors. The experiments demonstrated high effectiveness in both identifying postures and analyzing canine movements in diverse outdoor environments. Crucially, the method proved robust across varied lighting conditions, angles, and breeds.
One of the practical outcomes of this research is its implementation as a mobile application capable of real-time animal tracking. This enhances its applicability in welfare monitoring, working dog performance evaluation, wildlife observation, and consumer-oriented pet technology. By enabling accurate, markerless pose recognition, the study contributes to advancing accessible and noninvasive tools for canine behavior analysis.
Source: Srinivasan, R., Maskeliūnas, R., & Damaševičius, R. (2021). Markerless Dog Pose Recognition in the Wild Using ResNet Deep Learning Model. De Computis. Published December 24, 2021.
