Deadlift Pose Estimation: Landmarks, Camera, Rep Logic

This page is the deadlift layer under how rep counting works and how form feedback works: landmarks, camera placement, rep logic and what the three 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 deadlifts among 12 exercises with angle-threshold state machines on MediaPipe landmarks. Their test set held 231 deadlift videos: 160 they recorded, 47 from public pose datasets and 24 from competition and online footage.
• Mercadal-Baudart and colleagues (Heliyon, 2024) recorded four female and four male athletes doing squats, deadlifts and jumps in a marker-based motion-capture lab, with phones placed around them, plus four male athletes in a second collection. They trained a monocular 3D model to turn 17 detected 2D keypoints into a 33-point skeleton that includes the spine and pelvis.
• Shoaib and colleagues (Scientific Reports, 2026) classified deadlift form as good or bad from 377 single-repetition clips of three male lifters aged 19 to 25, filmed with two smartphones from five angles, using YOLOv11-Pose keypoints and graph neural networks. A certified trainer labelled each clip.

Landmarks and angles#

Sousa and colleagues' deadlift runs on two signals: a torso angle, leaning forward at the start and extended in the lift, and the elbow angle, which must stay straight in both states. Their pipeline computes knee, hip, elbow, shoulder and torso angles from MediaPipe's 33 landmarks.

The hip is the hard part. Mercadal-Baudart and colleagues point out that common skeletons such as OpenPose's have a single keypoint at each hip, which is insufficient for clinically meaningful hip joint angles because those depend on the orientation of the pelvis and femur, and that simplified torsos are inadequate for pelvic angle and flexion at different levels of the spine. Their skeleton adds the toes, pelvis points (left and right PSIS and iliac crest), four spinal points (L2, T10, T4 and C7), the sternum and the clavicle. They define hip flexion as the angle between the lower spine (L2 to hip) and the thigh, with knee and hip flexion set to 0° in quiet standing. In our reading, that is the opposite convention to a three-point interior angle that reads near 180° for a straight leg, so check which convention a borrowed threshold assumes.

Shoaib and colleagues used the 17 keypoints of the COCO standard and state that projected 2D keypoints cannot directly recover depth-sensitive quantities such as the 3D bar path or spinal curvature. On this site's keypoint maps, COCO 17 and BlazePose 33 have no spine or mid-hip point, OpenPose BODY_25 and Halpe26 add a hip centre and a neck, and Apple Vision 3D has a single spine joint and a root.

Where to put the camera#

Mercadal-Baudart and colleagues recorded with phones placed around the athlete at a radius of about 3 m, both on the floor and on tripods at about 1.5 m height, to simulate where physiotherapists and coaches would record. Their model was trained with virtual cameras that projected the motion-capture data into 2D from a range of possible camera positions, and they call for real-world testing of lighting, clothing and camera positioning.

Shoaib and colleagues filmed one sagittal or side view and four frontal or diagonal views, four indoors and one outdoors, to add viewpoint, lighting and background variability rather than to prescribe a set-up. They did not standardise camera height or lifter-to-camera distance, which they list as a limitation.

Sousa and colleagues noticed occlusion of key body parts particularly in bending exercises such as squats and deadlifts, more pronounced at greater distances. No study we read compared deadlift accuracy by camera view, so test the placements your users will actually use.

A rep state machine for deadlifts#

Sousa and colleagues divide the deadlift into two states: a starting position with the torso leaning forward below one threshold and the elbows straight, and a lifting state triggered when the torso extends above a higher threshold with the elbows still straight. Their counter increases on completion of a cycle. In our reading, requiring the torso to drop back below the lower threshold before the next lift can count is the hysteresis described in how rep counting works, and the straight-elbow condition keeps an arm movement from passing as a lift.

Shoaib and colleagues describe the deadlift as predictable phases, setup, eccentric lowering, concentric pull and lockout, with the setup and lockout relatively static and the concentric pull carrying the dynamic moments. Their trainer marked six checkpoints by bar position: bar on the ground, below the knees, at knee level, above the knees, mid-femur and final standing. A pose model does not see the bar; in our reading, the wrists, which hold it, are the nearest proxy, so wrist height against the knees marks the knee passage.

For dropped frames, Sousa and colleagues keep the last valid phase for typically 3 to 5 frames. Mercadal-Baudart and colleagues chose their 3D model partly because its temporal components help smooth jittery movement between frames.

Failure modes the studies report#

• Missed bottoms and thresholds. Sousa and colleagues' deadlift errors were seven angle-threshold failures and twelve missed bottom positions. Errors mostly meant missed repetitions rather than false positives, particularly in squat and deadlift videos recorded in uncontrolled environments.
• Occlusion at a distance. Key body parts were occluded in bending exercises, more so at greater distances.
• What 2D cannot see. Spinal curvature and the 3D bar path are not recoverable from projected 2D keypoints, and a single hip point cannot give pelvis orientation.
• New people and new conditions. Shoaib and colleagues list sensitivity to pose-estimation noise and viewpoint or occlusion effects, and with their frame selection, four of their six models lost accuracy significantly on an expanded test set that added 54 repetitions from 16 outside lifters.

Pose estimation accuracy covers the general causes and 2D vs 3D pose estimation what estimated depth can add.

What was measured, and why it will not transfer#

• Sousa and colleagues counted deadlifts with 95% accuracy, slightly below most exercises, which they attribute to pose estimation noise affecting the angle calculations.
• Mercadal-Baudart and colleagues' root mean square angle errors were within 10° for all exercises for shin angle, knee varus/valgus and flexion, hip flexion and pelvic tilt, trunk angle, spinal flexion and rib flare, with larger errors, still within 15°, for shoulder flexion and ASIS asymmetry in some exercises. They attribute the high accuracy to training on ground truth that contained these same exercises, and found that prediction accuracy reduces quickly for movements the model was not validated on.
• Shoaib and colleagues' proposed model reached 89.5% accuracy on held-out videos of their three lifters. With a lifter held out entirely, accuracy ranged from about 50 to 74%, so the authors read their main result as subject-dependent rather than full cross-person generalization. The full pipeline took about 0.8 s per repetition on a desktop GPU, not a phone.

None of this transfers on its own to another model, camera position or group of users. Shoaib and colleagues describe their system as 2D projection-based form screening rather than 3D biomechanical measurement. Testing a rep counting algorithm covers building your own test set.