Extended Data Table 2 Quantitative evaluation on 3D reconstruction

From: Controlling diverse robots by inferring Jacobian fields with deep networks

  1. Using testing images from all camera viewpoints, we measure depth prediction errors in millimeters and optical flow prediction errors in pixels on a 640 × 480 image grid. The depth errors are computed as the L2 distance between the ground truth depth value measured by the Intel Realsense D415 RGB-D cameras and the predicted depth value averaged over all pixel locations in the image. The optical flow errors are computed as the L2 distance between the ground truth measured by the point tracker58 and the prediction averaged over all pixel locations in the image.
Back to article page