Fig. 4: Landmark detection on novel setups using test-time scale optimization.

a Sample images from a challenge dataset, collected from various experimental setups. b Pixel error when running a trained model on a video clip from the challenge set. Each row represents a different spatial scale factor applied to the video clip prior to landmark detection. c Mean prediction error in pixels versus scale factor. The model exhibits a preferred scale. d Two schemes for test-time scale optimization: frame-level (solid) and clip-level (dashed). In frame-level optimization, each frame receives its own scale factor. In clip-level, all frames share the same scale factor. e Precision-recall curves (left; numbers indicate area under the curve) and pixel error quartile plots (right; center lines and numbers indicate median, boxes indicate 25th and 75th percentiles) for the different types of scale optimization compared to performance with no scale optimization as well as small-scale regime networks from Fig. 3g, h (n = 612 test images). Clip-level scale optimization achieves the highest performance. Source data are provided as a Source Data file.