Mandora localizes joint positions with high accuracy. a-b, We develop a deep convolutional neural network model for detecting locations of multiple joints from finger, wrist, and toe. Instead of using the pixel as the unit, we measure the difference between prediction and ground truth using a normalized distance so that images with different sizes are comparable. c-e, The predictive performances of locating joints from the finger, wrist, and toe are shown as box plots. The horizontal dashed red line represents the cutoff, a normalized distance of 0.02. The numbers represent the percentage of samples with a smaller distance than the cutoff. f-g, Two examples are shown with normalized distances of 0.01 and 0.02, respectively.