Skip to main content
Translation updated
#3d-human-pose-estimation#transformer#graph-neural-network#motion-reconstruction

Coordinate frames for comparing 3D human pose estimates

Compare the inputs and outputs of MotionAGFormer, HybrIK and TRAM across 2D-to-3D lifting, mesh recovery and world-space motion recovery.

“3D human pose estimation” covers several tasks. Lifting 2D joints into camera-relative 3D, recovering an SMPL mesh from RGB, and recovering a person’s world-space trajectory with a moving camera all require different inputs, outputs, and metrics. This article groups the papers by the question each model answers.

Sources checked on August 25, 2026. Compare numerical results within the protocols defined by each paper. The links lead to CVF, ECVA, DOI records, and the authors’ repositories where available.

#Three estimation tasks

TaskTypical inputOutput coordinatesTypical error metric
2D-to-3D pose lifting2D joints per frameRoot-relative 3D jointsMPJPE, PA-MPJPE
Human mesh recoveryRGB image/videocamera-relative joints + body meshMPJPE, PA-MPJPE, PVE
Global motion recoverymoving-camera videoworld-space trajectory + local body motionglobal position/orientation/trajectory error

The first task can omit global translation and scale. The second includes body surface and shape, though position in the world can remain ambiguous. The third also requires camera motion and metric scale.

#MotionAGFormer lifts a sequence of 2D joints into 3D

MotionAGFormer estimates 3D joints from a time series of 2D joints. A Transformer stream models global joint relationships. A GCNFormer stream models local connections in the skeleton graph. The model adaptively fuses the two representations.

Check the metric and protocol separately for each dataset. End-to-end accuracy also depends heavily on whether the 2D inputs are ground-truth joints or detector predictions.

#HybrIK combines joint accuracy with mesh structure

HybrIK proposes hybrid inverse kinematics to map 3D joints to body-part rotations. It solves swing rotation analytically from the joints and learns twist rotation from visual cues. The result is a mesh that represents pose and body shape.

PVE and MPJPE measure error on different outputs. A pose-only model that produces no mesh cannot be compared directly on PVE.

#TRAM recovers world-space motion from a moving camera

TRAM appeared at ECCV 2024, correcting the 2025 date given in an earlier version of this article. The authors adapt SLAM to handle moving people and recover camera motion and scene scale. They combine these estimates with VIMO’s local body motion to reconstruct the world-space trajectory.

Camera-relative MPJPE captures only part of TRAM’s output. Evaluation also covers drift that makes a walking person appear to walk in place, separation of camera and human motion, and metric scale.

#Why a monocular image admits multiple 3D poses

Different 3D poses can produce similar 2D projections. Occlusion, foreshortening, and left–right symmetry make them harder to distinguish. A model narrows the possible solutions using learned body priors and temporal smoothness. Details absent from the image remain inferred.

For robot safety or medical measurements, consider confidence, uncertainty, and out-of-distribution signals alongside the estimated coordinates.

#What to record in a fair comparison

input | detector | temporal window | output representation | coordinate frame
training datasets | test protocol | metric | params/MACs | FPS/hardware
occlusion handling | multi-person tracking | license | code/model version

“RGB input” leaves several choices unspecified. Record the 2D detector, bounding-box crop, test-time augmentation, and whether the evaluation uses root alignment or Procrustes alignment.

#Choosing an approach for a robotics project

  • If only joint directions matter, start by measuring the latency of a pose-lifting model.
  • If collision distances and body volume matter, consider calibrated depth or a mesh with uncertainty estimates.
  • To recover world-space human trajectories from a moving camera, choose an approach that estimates camera pose and scale together.
  • For a safety stop, the ability to detect failure under occlusion or loss of view matters more than mean MPJPE.

#Reproduction steps

  1. Pin the commit and weight checksum from the paper’s official repository.
  2. Check the dataset license and evaluation protocol.
  3. Report results with supplied 2D keypoints separately from those with your detector.
  4. Keep camera-relative and world-space metrics separate.
  5. On the same video, measure latency, peak VRAM, and throughput as the number of people increases.
  6. Document failures under occlusion, motion blur, and camera motion.

As models change, the same questions still make comparisons meaningful. Specify the input, representation, coordinate system, alignment, and evaluation protocol before interpreting a ranking.

Connect