Research map / 3D vision / 3D generation
TULIP: Multi-camera 3D Precision Assessment of Parkinson's Disease
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces TULIP, a markerless 6-camera video dataset of 25 UPDRS motor exam activities from 15 subjects with clinician ratings, and benchmarks UPDRS classifiers on 3D vs 2D pose features for finger tapping and gait.
Problem
Parkinson's disease motor assessment via UPDRS is coarse and subjective, and there are no benchmark multi-view video datasets for PD motor exams that enable markerless 3D kinematic analysis.
Approach
A multi-camera markerless video dataset covering nearly all UPDRS Part III motor activities, with multi-clinician labels, plus baseline classifiers showing 3D pose features outperform 2D features for predicting UPDRS scores.
Key results
- 3D-feature MLP reached 0.69 weighted F1 on finger-tapping UPDRS score prediction and 0.87 F1 on PD vs healthy diagnosis.
- 3D-feature Random Forest reached 0.72 F1 on gait UPDRS score prediction, outperforming two clinicians (0.67-0.73 F1 range for clinicians overall).
- Models using 3D features consistently outperformed 2D-feature models across both activities.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 3D rather than 2D features | UPDRS classifiers on TULIP finger-tapping recordings · Index finger tapping | weighted F1 | Helped |
| use 3D rather than 2D features | UPDRS classifiers on TULIP gait recordings · Gait | weighted F1 | Helped |
| use a ResNet50 CNN on 3D pose sequences instead of the 3D-feature MLP | UPDRS classification on TULIP finger-tapping recordings · Index finger tapping | weighted F1 | Hurt |
| use a previously published Random Forest approach with a smaller set of 3D features instead of the proposed gait model | PD/healthy classification on TULIP gait recordings · Gait | F1 | Helped |
Surprising findings
- A ResNet50 trained directly on 3D pose sequences performed worse than the 3D feature MLP, possibly due to the small dataset size.
- Inter-clinician agreement on some UPDRS components was very low (e.g., kinetic tremor ICC of 0.1 to 0.4).
Limitations
- Small cohort of 15 subjects with limited demographic diversity.
- Baselines use only finger tapping and gait out of 25 recorded activities.
- Recorded in a large room that does not reflect a typical clinical exam environment.
Experimental setup
Datasets: TULIP
Benchmarks: TULIP
Metrics: weighted F1 F1
Baselines: ResNet50 on 3D pose sequences LightGBM on 2D features SVM Random Forest AdaBoost XGBoost LightGBM MLP
Builds on: UPDRS MediaPipe MMPose
Closest papers by meaning
- Care-PD: A Multi-Site Anonymized Clinical Dataset for Parkinson’s Disease Gait Assessment (NeurIPS 2025 · similarity 0.90)
- Accessible, At-Home Detection of Parkinson’s Disease via Multi-Task Video Analysis (AAAI 2025 · similarity 0.89)
- Breaking Data Silos in Parkinson’s Disease Diagnosis: An Adaptive Federated Learning Approach for Privacy-Preserving Facial Expression Analysis (AAAI 2025 · similarity 0.82)
- DDx-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs (NeurIPS 2026 · similarity 0.80)
- EMAD: Evidence-Centric Grounded Multimodal Diagnosis for Alzheimer's Disease (CVPR 2026 · similarity 0.79)
- VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.