Bethell, E. J., Khan, W., Hussain, A. 2022. A deep transfer learning model for head pose estimation in rhesus macaques during cognitive tasks: Towards a nonrestraint noninvasive 3Rs approach. Applied Animal Behaviour Science 255, 105708.
Head orientation is a measure of attention used in behavioral psychological research with non-human primates. It is used across a broad range of disciplines and settings, from the field to the laboratory. Field methods are time consuming with risk of coding bias and visibility issues with free-ranging animals. Laboratory methods may require restraint and use of invasive procedures. Automated systems to measure head orientation in unrestrained animals, that are robust to partial occlusion of the head, would improve coding efficiency and accuracy and provide 3Rs animal welfare benefits. We present a free-to-use deep transfer learning model for non-invasive head pose estimation in unrestrained Macaca mulatta taking part in cognitive experiments. Monkeys housed in social groups were filmed viewing two conspecific face stimuli presented on either side of a video camera. Video frames were manually annotated for three head positions relative to the video camera: ‘left’, ‘center’ and ‘right’. The dataset (total = 8135 images from 26 monkeys) was partitioned into training and testing datasets using a leave-k-out strategy, so that 70% of the images were used in training and 30% were used in testing. We used the VGG16, VGG19, InceptionV3 and Resnet50 as base models to train the proposed head pose classifier. We achieved model accuracy up to 93 %. The head pose estimation model presented here will be of use across contexts ranging from field-based playback experiments to assessment of welfare in zoo and clinical veterinary settings and refinement of neuroscience research practices. Model code with instructions is provided.