Rep counter playground: tune the thresholds on your own video

How the counter decides

It is the same state machine as the 30-second rep challenge, with every constant exposed. For each frame:

1. One angle. The chosen joint’s angle at its middle point: knee (hip, knee, ankle), hip (shoulder, hip, knee), elbow (shoulder, elbow, wrist), shoulder (hip, shoulder, elbow). A side is read only if all three points are at least the minimum visibility the model reports; the side setting picks left, right, the average of the visible sides, or the more visible one.
2. Smoothing. An exponential moving average with a time constant you set (50 ms to start), so it lags by about the same time at any frame rate. The faint line on the chart is the unsmoothed angle.
3. Two zones and a gap. For a movement that starts with the joint open (a squat, a push-up, a curl), the rest zone is above the top angle and the active zone below the bottom angle. The counter waits until the angle first holds in the rest zone, then needs the active zone, then the rest zone again: that return is one rep. An angle between the two thresholds changes nothing, which is why a wider gap rejects more jitter and misses more shallow reps.
4. Holding and pacing. A zone counts only after the angle stays in it for the minimum number of consecutive frames (2 to start), and a cycle completed faster than the minimum cycle time (600 ms to start) is drawn as a dotted tick and rejected, not counted.

The page’s own starting values (top 160°, bottom 100° on the knee) are the challenge’s squat rule and are not calibrated for any exercise or camera. Every model frame’s angles are kept in memory, so changing a threshold recounts the whole recording instantly without running the model again. That is the use of the page: find the values at which the count matches what you see, on footage shot the way your users will shoot it.

Starting from published rules

The preset list loads the thresholds that licensed open-source exercise trackers ship, from the open exercise form-rules dataset: 4 rules from 3 projects (Ultralytics AIGym (workouts monitoring solution), AGPL-3.0; Exercise-Correction (Ngô Hồng Quốc Bảo), MIT; PythonExcercisePoseEstimation (thphuc05), MIT), each number read from the project’s code on 2026-11-05 with the line it came from. A preset sets the joint, side, direction and the two angles; the frames, cycle time and smoothing stay this page’s, because those projects count differently (some on the way down, some on the way back) and none states them in the same terms. Their thresholds are the authors’ choices for their own setups, not biomechanical standards, and the sources compute angles in 2D image coordinates, which is why that is the default here.

The exported JSON names its schema, every setting and, for a preset, the dataset rows it came from, so a config in your repository says where its numbers started.

Reading the trace

• A rep is missing when its trough stops short of the bottom line, or when the angle never climbs back above the top line between reps. Move the line to the shape you see rather than the reverse.
• A rep is counted twice when noise crosses a line more than once. Widen the gap, raise the minimum frames or the smoothing, or raise the minimum cycle time.
• Gaps in the line are frames with no reading: no person found, or the joint’s points below the visibility floor. Side-on footage hides the far limb, which is what the side setting is for. Camera placement is covered in how rep counting works and improving pose detection accuracy.
• Image or world angles. Image angles are measured on the picture, with x scaled by the frame’s aspect ratio, so they shrink when a limb points at the camera. World angles use the model’s 3D landmarks in meters and depend less on where you stand, at the cost of the model’s depth estimate.