Static poses are classified by shape; dynamic gestures are classified by trajectory

What it is

A static pose (hand pose) is a shape on a frame: open palm, fist, a “V.” Time barely enters the model. A dynamic gesture is a trajectory: how position, orientation, and speed change. Both can be mid-air commands; they fail differently. Poses fail when fingers are occluded or the shape is far from the template. Trajectories fail on speed, amplitude, and slicing. Writing “open the hand” and “wave right” as one kind of “gesture” makes the recognizer, the lesson, and the evaluation share assumptions they should not share.

Why it happens

Pose recognition eats joint angles and fingertip layout on a single skeleton or depth frame; the classifier can be frame-independent. It does not need onset and offset, but it does need that frame's hand shape to be fully visible. Trajectory recognition eats a time series. Hidden Markov models, dynamic time warping, or temporal networks align a path. They are less hungry for a perfect single-frame shape (fingers may blur in a wave) and more sensitive to time alignment: the same spatial path, twice as fast or twice as slow, inflates the distance. A pose can constrain a dynamic gesture as a keyframe, but a keyframe is not a trajectory. Teaching a wave with three hand-shape stills makes learners freeze on those stills while the recognizer wants the motion in between.

Studying it

Split the vocabulary into “shape only” and “must travel a path,” and report confusion matrices separately rather than one recognition rate. For poses, factors are viewpoint, occlusion, and camera distance. For dynamics, factors are speed, amplitude, and segmentation policy. Crossing conditions diagnose: hold the correct shape and do not move—does the dynamic recognizer still fire? Travel the path with a wrong shape the whole way—does a pose gate stop it? Datasets that label still photos and video clips with one tag set make later papers unable to say which model class the errors came from.

Where it stops holding

Some commands mix: pose then travel (pinch then drag). Mixed items must be written in parts—which stretch is shape, which is trajectory—or debugging cannot tell whether to add keypoints or time alignment. In signed language, fingerspelling leans static and lexical signs lean dynamic; moving them onto an interface vocabulary does not license one sensor setting. RGB-only, no depth, collapses static poses from the side or back into similar silhouettes; trajectories in a plane perpendicular to the camera flatten too. A headset looking at its own hands has a stable view, so poses are cheaper. A room camera watching passers-by has more stable trajectories and more brittle poses.

Applying it

• Tag every vocabulary item: pose, trajectory, or pose-plus-trajectory. Split teaching, recognizers, and test sets by type. Do not share one “gesture accuracy.”
• Check poses under viewpoint and occlusion; check trajectories under fast/slow and segmentation. Do not patch one class of error with data from the other.
• Write mixed commands as a state machine: enter a shape, hold or move, leave—not as one end-to-end label.

Related

• Same group: C4.04.2 Static poses require holding and impose continuous muscle load · C4.04.3 Dynamic gestures vary more across people in speed and amplitude
• Adjacent: C4.03 Gesture onset and offset detection · C4.09 Finger count as input
• Search: static hand pose · dynamic gesture · trajectory recognition