For pet owners who want answers beyond “looks a bit stiff,” and for veterinarians who want objective kinematics beside clinical judgment. Research-grade pose estimation, four-plane protocol, and the honesty to suppress numbers it cannot stand behind.
A skilled clinician's eye remains essential. The gap is repeatable measurement: mild, bilateral, or well-compensated lameness is easy to under-call, hard to score the same way twice, and difficult to track across weeks of rehab.
A few walk passes. An impression of grade. Mild or bilateral patterns — and day-to-day drift between observers — leave owners and referring vets without a shared number.
Quadrupeds redistribute load. What looks “a bit stiff” can be measurable asymmetry in stride, stance, head bob, pelvic drop, or sit-to-stand weight-shift.
Force plates are gold-standard for load — and largely confined to research centres. Day-to-day rehab still lacks objective, clinic-native kinematics.
Real RehabVet clinic captures with the skeleton overlay our pipeline draws — switch planes, inspect joint groups, and read the metrics each view unlocks. Videos autoplay muted and loop.
Interactive pose stream
Tap a region — see what we measure
Illustrative walk cycle using the same keypoint layout our model outputs. The video on the left is a real RehabVet session with the live overlay.
Why this plane
The workhorse view for joint angles and timing.
Selected: Forelimbs — Elbow / carpus ROM · tracking
A side-view exam is powerful — and blind to tracking and pelvic drop. Frontal and rear views, plus sit-to-stand, close the gaps that a few walk-by eyeball passes leave open.
The workhorse view for joint angles and timing.
What a side camera cannot see — tracking and pelvic drop.
Load transfer under effort — often the first place pain shows.
Each number maps to a kinematic or spatiotemporal concept from our analytics pipeline. Toggle the voice that fits you.
Left–right symmetry index on mean stride time (forelimb and hindlimb). Higher % = greater temporal asymmetry.
Symmetry index on mean stance time per paw. Complements stride timing when load-sharing is uneven.
Per-joint angular ROM from pose trajectories (stifle, hock, elbow, carpus). Side-view kinematics, not kinetics.
Lateral: vertical head oscillation asymmetry. Frontal: lateral head bob amplitude and L/R asymmetry when coverage allows.
Pelvic asymmetry from hip trajectories (lateral) and pelvic drop / hock alignment on rear retreat.
Bilateral forelimb tracking symmetry and related frontal measures — independently gated when coverage is sparse.
Transition score (0–5), weight-shift asymmetry, head compensation, attempt count, and transition time.
Coverage × data-quality factor (e.g. stride CV / stride count). Metric groups can be null when below coverage thresholds.
Built on production clinic footage: YOLO26 for subject detection, DeepLabCut SuperAnimal for pose, and RehabVet analytics with reliability gating — clinic adaptations of DeepLabCut and YOLO-family research lineages detailed below. Impressive where deserved — honest where evidence stops.
Ultralytics YOLO26 (clinic default: yolo26l) runs first. It finds the dog bounding box frame-by-frame (with optional tracking), so pose estimation is cropped to the patient rather than a handler or background clutter.
Pose estimation uses DeepLabCut’s SuperAnimal-Quadruped model with an HRNet-W32 backbone — a research-grade, markerless quadruped pose network that outputs 39 standardised body parts (eyes, ears, spine chain, limbs, paws, digits, withers, throat, chin, and tail landmarks).
Pose estimation builds on DeepLabCut research originally developed at Harvard University. Subject detection builds on the YOLO family of methods originally introduced at the University of Washington — adapted here with clinic-tuned models and RehabVet analytics. SuperAnimal enables multi-species quadruped pose without per-clinic marker training; analytics and reliability gating are clinic-owned — developed against real RehabVet sessions, not a consumer phone app.
We do not publish a peer-reviewed validation study claiming force-plate equivalence. What we do publish: quality-weighted confidence (coverage × data-quality), independent frontal metric-group gating (e.g. forelimb vs head/trunk at ~15% frame coverage), and illustrative clinical cases with real numbers. Kinematics ≠ load — complementary to force plates, not a replacement.
Chart weight maps the patient to a size class (toy → giant) that calibrates lameness-grade thresholds. Toys can grade slightly more conservatively for the same raw asymmetry — measurement noise is relatively larger at small scale. Breed adds mild morphotype context (for example chondrodystrophic / short-legged conformation). Weight is primary for size; breed supplements conformation. Missing chart weight defaults to medium — clinics should record weight for better calibration. This is score calibration, not better DeepLabCut keypoints; film quality and clinical judgment still matter.
A multi-plane clinical capture — lateral walk, frontal approach, rear retreat, and sit-to-stand — filmed under controlled clinic conditions. No markers. No wearables. No force-plate lab.
YOLO26 locates the dog in every frame so pose estimation stays locked on the patient — even when a handler is in shot. Dog-preferring selection reduces person/background confusion.
DeepLabCut 3.0 SuperAnimal-Quadruped (HRNet-W32) estimates 39 anatomical keypoints — nose to paws — with a per-joint confidence score on each frame.
Stride timing, stance symmetry, joint ROM, head bob, pelvic drop, tracking symmetry, and sit-to-stand weight-shift — computed from the pose stream, not guessed from a glance.
If a view or metric group is too sparse, we suppress it. Frontal forelimb and head/trunk groups are gated independently. Numbers that reach the report are ones the system is willing to stand behind.
A lameness score (0–5), affected-limb localisation, owner-facing language, and a full clinical narrative for the veterinary team — always verify against the video.
Four July 2026 clinic captures — illustrative of methodology and gating, not a controlled trial.
3.2
Lameness / 5.0
Left Rear90% confidence · 767 frames
Lateral walk overall 3.17/5 localised to left rear (hindlimb 3.17 vs forelimb 2.62). Hindlimb stride asymmetry 40.2%; stance asymmetry 22.7%; head-bob asymmetry 8.7%. Sit-to-stand overall 4.29/5 (confidence 0.9) with weight-shift asymmetry 155.7% — corroborating the same limb.
3.3
Lameness / 5.0
Right Front90% confidence · 954 frames
Lateral walk overall 3.28/5 localised to right front (confidence 0.9, forelimb score 3.28). Head-bob asymmetry 16.3%; pelvic asymmetry low (3.5%). Sit-to-stand overall 3.9/5 (confidence 0.9) with marked head compensation (121.7 px) and 3.2 s transition time.
3.3
Lameness / 5.0
Right Rear90% confidence · 703 frames
Lateral walk overall 3.30/5 localised to right rear (confidence 0.9). Head-bob asymmetry 19.1%; pelvic asymmetry 29.4%; hindlimb stance asymmetry 32.9%. Frontal forelimb tracking produced no usable frames and was gated. Sit-to-stand 3.46/5 in 1.8 s on first attempt.
3.3
Lameness / 5.0
Left Rear90% confidence · 546 frames
Lateral walk overall 3.31/5 localised to left rear (confidence 0.9; hindlimb 3.31 vs forelimb 2.38). Hindlimb stride asymmetry 38.9%; stance asymmetry 17.4%; head-bob asymmetry 7.4%. Sit-to-stand overall 3.60/5 with weight-shift asymmetry 183.8%.
| Visual gait exam | Force-plate lab | RehabVet AI | |
|---|---|---|---|
| What it measures | Subjective impression of limp / asymmetry | Ground reaction force (load) | Multi-plane kinematics + timing + sit-to-stand |
| How the call is made | Clinician eyeballs a few walk passes | Instrumented walkway / force plate | 39-keypoint pose model + gait analytics |
| Repeatability | Varies by observer, lighting, and day | High — under lab conditions | Standardised protocol · confidence-scored |
| Reliability posture | Always forms an impression | Gold-standard kinetics | Suppresses unreliable metric groups |
| Where it lives | Every consult — but unquantified | Research / specialty centres | Live rehab clinic · R&D in production care |
RehabVet is Singapore's first — and one of its most comprehensive — animal rehabilitation clinics. This gait intelligence is being pioneered in the same rooms where we treat IVDD, cruciate disease, and arthritis every week: eight years of hydrotherapy, physiotherapy, TCVM, and outcomes care informing the science.
Read the June 2026 Impact ReportObjective movement data should be as ordinary in rehab as a thermometer — and as honest about its limits.
Dr Sara Lam · Founder, RehabVet
Scores and narratives support clinical reasoning. They do not replace physical examination, imaging, or a veterinarian’s diagnosis.
Blur, occlusion, short clips, extreme angles, or handler interference reduce keypoint coverage. Poor input yields lower confidence or suppressed metrics — by design.
Frontal forelimb tracking is often the hardest view. When coverage falls below threshold, those metrics are null (see Haru). Absence of a number is not proof of normality.
We measure movement patterns from video. Force plates measure load. Use both when you need the full biomechanical picture.
The four sessions on this page are real clinic captures with reported metrics. They are not a controlled peer-reviewed validation cohort. Methodology and honesty come first.
The clinical protocol and “world’s first” claim are dog-scoped as of July 2026. Cat support is on the roadmap; SuperAnimal itself is quadruped-capable.
For pet owners
For veterinarians
Technology & accuracy
Data & limits
Book a clinical assessment at RehabVet — or write to us for referring-clinic coordination and research collaboration.
RehabVet AI Pet Mobility Assessment supports — and does not replace — veterinary diagnosis. Results depend on footage quality. Always verify findings against the video and clinical examination.
