Overhead Press Pose Estimation: Landmarks, Camera, Rep Logic

This page is the overhead press layer under how rep counting works and how form feedback works: landmarks, camera placement, rep logic and what the studies we read on October 18, 2026 measured. Each number is one study's result under that study's conditions, so use it as a starting point to calibrate for your camera, model and users. This is engineering guidance for building a tracker, not exercise or medical advice.

The studies behind this page#

• Sousa and colleagues (Journal of Functional Morphology and Kinesiology, 2026) counted shoulder presses among 12 exercises with angle-threshold state machines on MediaPipe landmarks. Their test set held 254 shoulder press videos: 160 they recorded, 45 from public pose datasets and 49 from competition and online footage.
• Hwang and colleagues (Sensors, 2023) had 30 healthy adults with less than three months of exercise experience do five seated upper-body exercises, shoulder presses among them, with dumbbells and resistance bands, and counted repetitions from OpenPose BODY_25 keypoints.

Landmarks and angles#

Sousa and colleagues define the shoulder press with the elbow and shoulder angles: a starting position with the weights at shoulder level and both angles generally below a threshold, a pressing phase in which the arms extend upward and both angles rise above a higher threshold, and a lowering phase in which the angles return below. During the pressing phase their system also checks, from the relative position of keypoints, that the hands and elbows rise above the head. Hwang and colleagues count from the vertical position of the left hand instead, after dividing coordinates by torso length to remove the effect of camera distance and body size.

Every keypoint scheme on this site has shoulders, elbows and wrists. For the hands-above-the-head check you also need a head reference: COCO 17 and BlazePose 33 have the nose, eyes and ears, and Apple Vision 3D has a top-of-head joint. In our reading, the shoulder angle is the weaker of the two signals: in a lunge study we read, OpenPose's shoulder flexion angle agreed only poorly to moderately with marker-based capture from almost every camera view, so give a shoulder-angle threshold more margin than an elbow one.

Where to put the camera#

Hwang and colleagues placed their camera 1.5 m in front of the seated participant at a height of 1.2 m. Sousa and colleagues did not fix one position: part of their test set came from home environments, commercial gyms and competition venues with different lighting, backgrounds and camera angles, and they note that their pose model's performance can degrade when camera angles deviate from frontal or lateral views. In our reading, a front view is the natural start for a press, because the hands-above-the-head check needs both arms and the head in the image.

A rep state machine for the overhead press#

Sousa and colleagues' three states give the counter its hysteresis: the angles must fall below the lower threshold before another press above the higher one can count, as described in how rep counting works. Their phase persistence keeps the last valid phase for typically 3 to 5 frames when no phase matches. The same module composes into other exercises: a thruster is a squat module followed by a shoulder press module, and an overhead squat combines a squat with the shoulder press's second phase, the arms extended.

Hwang and colleagues count the zero-crossings of the hand-height signal after a 15-sample moving median filter, which they use to cope with occlusions shorter than about 0.25 s, and after removing the signal's average.

Failure modes the studies report#

• Hands behind the head. In a related overhead movement, the overhead triceps extension, Hwang and colleagues found the hand usually behind the head, causing an obstruction their median filter could not fully resolve, and counting accuracy was low.
• Short or uneven sets. Their periodicity-based count tended to be overestimated in sets with few repetitions, where the timing of each repetition varied, and the method is vulnerable to non-repeating, nonlinear movements with heavy weights.
• Occlusion and fast transitions. Sousa and colleagues attribute their errors mainly to temporary occlusions, rapid transitions and suboptimal camera viewpoints, which mostly produced missed repetitions rather than false positives.

Pose estimation accuracy covers the general causes in more depth.

What was measured, and why it will not transfer#

• Hwang and colleagues counted shoulder presses, chest presses and seated rows accurately, with small average absolute errors, while biceps curls and overhead triceps extensions had larger errors; across the five exercises, the mean absolute counting error ranged from 0.88 to 3.21 repetitions.
• Sousa and colleagues' per-exercise counting accuracies ranged from 95 to 100% across their 12 exercises, and thrusters, a squat followed by a shoulder press, reached perfect detection (100%).

The presses in these studies were seated in one and filmed in mixed settings in the other, and neither population is yours. Measure your own counter before you rely on it; testing a rep counting algorithm covers how.