Anchor refinement
A lattice of anchor priors sits over the frame; the matched prior's four corners regress outward along their offset vectors onto the true object extent and lock there.
A lattice of anchor priors sits over the frame; the matched prior's four corners regress outward along their offset vectors onto the true object extent and lock there.
A labelled sample steps through mirror, rotation and crop transforms with its annotation box carrying through each one, while a chip row marks the transform in force.
A confidence field spreads wide then contracts through its contour levels onto one grid cell as the argmax crosshair closes on the located point.
A predicted box slides onto the dashed ground truth while the shaded intersection grows and the ratio rule beneath climbs past the match threshold.
Detected joints pop in one after another and the limb links draw between them, assembling a loose keypoint set into a connected pose graph.
A barrelled calibration target unwinds onto the rectilinear reference as the fitted distortion model is applied and the residual rule collapses toward zero.
Four overlapping candidate boxes ranked on a score ladder are pruned from the lowest score upward until only the maximum-scoring detection survives.
A tracked box crosses a hatched occluder, coasts through it as a dashed prediction carrying the same identity chip, and resumes as a solid detection on the far side.
A lattice of flow vectors anchored at their sample points extends to full magnitude as a displacement wave crosses the frame, then relaxes behind it.
A captured page sits sheared off axis with its four detected corners marked; the warp unwinds until page and corners land on the target rectangle.
A tracker runs a target along its track while the dashed prediction repeatedly leaps ahead of the measured centroid and is pulled back by each correction.
Inside a held detection box a scan head descends and a mask follows it, promoting a faint object contour into a filled per-pixel segmentation row by row.
A search window walks back along the epipolar line in the right view, dwelling on each candidate until it settles on the reference patch and the offset dimension reads out.
A kernel window steps across the input grid two cells at a time and returns to the next row, writing one activation into the smaller feature map at every stop.
A scan head crosses a detected text-line box and the character cells beneath it resolve left to right into a decoded sequence.