Impact of Patient Orientation in Single- and Multi-View Camera Environments for AI-based Rehabilitation Monitoring

What it claims

Camera viewpoint materially changes how well a personalised, similarity-based rehabilitation monitor separates correct from faulty repetitions. The paper ships REHAB26-ViewAngles, a dataset of correct and erroneous exercise repetitions recorded across many patient orientations, and a separability metric scored inside a specific few-shot monitoring pipeline. Headline numbers:

• The best single 2D view beats the standard 0° frontal view by 16.9% in separability.
• The best single 2D view often beats single-camera 3D estimation.
• Adding a second view gains up to 13.1% more.

The design targets one or two uncalibrated smartphones. The authors treat more than two cameras as unrealistic.

How it works

The results are tied to one evaluation pipeline, adopted from the authors' earlier work. It is not a trained classifier:

1. A physiotherapist records a few of the patient's repetitions in a mobile app and labels each one as correct or not.
2. Every recording is processed the same way: - pose estimation - automatic segmentation into repetitions - interpretable features describing relationships between body segments
3. The labelled-correct repetitions provide two things: - per-feature, per-time confidence intervals - a single reference repetition
4. Each home repetition is temporally aligned to the reference. Its feature values are checked frame by frame against the intervals. Weights based on how discriminative each feature is produce an overall quality score and point to the main errors.

flowchart LR
  PT[PT-labelled correct reps] --> CI[Confidence intervals]
  PT --> REF[Reference rep]
  HOME[Home rep] --> AL[Temporal alignment]
  REF --> AL
  AL --> CMP[Compare to intervals]
  CI --> CMP
  CMP --> W[Weighted score + top errors]

Pose sources compared:

• Single-camera monocular 2D and 3D.
• Two orthogonal 2D views combined.
• Geometric triangulation.
• Weighted fusion of independently estimated 3D sequences.
• RUMPL, a cross-view transformer that maps 2D keypoints to 3D ray vectors, trained here on two synchronised cameras.

The related-work section names MediaPipe Pose as the most widely adopted on-device estimator. It produces 2D and 3D output from a BlazePose detector plus the GHUM body model. The fetched text does not confirm that it is the estimator used in the experiments.

Data acquisition:

• Three fixed cameras at θ = −90°, −45° and 0°.
• The participant rotates their standing angle α relative to the 0° camera.
• Effective viewpoint is VP = α + θ, so each pose of the participant yields three viewpoints at once.
• At each orientation the participant performs 10 continuous repetitions: 5 correct and 5 erroneous.

Limits

• Fetched text is truncated. It stops in the dataset section. It does not include:
• dataset size (participants, exercise types, number of orientations)
• the separability metric's definition
• per-exercise optimal angles or result tables
• any statement on dataset availability or licence

Check the full PDF before relying on the headline percentages for a specific exercise.

• Results are relative to this pipeline. The gains are measured with the personalised similarity scorer. They may not carry over to end-to-end learned quality classifiers.
• RUMPL needs calibration. The authors note it tolerates arbitrary camera placement but still needs valid calibration to orient its rays. That is in tension with the uncalibrated home setting.
• Errors are scripted. Erroneous repetitions are performed on purpose in blocks of 5 within a continuous set. They are not spontaneous patient errors.
• The paper is a v1 submitted for review on 7 Aug 2026.