Pixels never occupied by keypoints

[SylwestrakLab]

Our tracking is working well at the macro scale. However, when I plotted one keypoint across the whole experiment (in this case, elevated plus maze), I noticed that when I zooming into an example frame with the inference points overlaid from the whole experiment, only every other row/column of pixels is occupied by keypoints. Is this an artifact of downsampling? We are not downsampling our images at any point, and I had thought the network was trained on a downsampled image for the coarse identification of the subject, but then used a full image to identify the keypoints. Thanks in advance for your help!