Deep Learning / Object Detection / 2D Pose Estimation / Computer Vision / Sports Biomechanics / 2D Center of Mass Estimation / Temporal Filtering / Marker-less Analysis of Springboard Diving

DivingFy: Marker-less Estimation of Diver's Center of Mass Trajectory in Springboard Diving Using Deep Learning and a Single Fixed Camera

Resources will be shared when the paper is accepted

Overview

DivingFy is a marker-less system for estimating the 2D center of mass (CoM) trajectory of divers projected onto the sagittal plane from monocular video captured with a single fixed camera. The system integrates Faster R-CNN for multi-object detection (diver, springboard, splash), HRNet for 2D pose estimation, Hampel filter for outlier removal, Butterworth low-pass filter for temporal smoothing, and Winter's 2009 anthropometric model for biomechanically grounded 2D CoM calculation. DivingFy achieves sub-centimeter precision (median 0.47 cm) in uncontrolled competition environments without specialized equipment.

Abstract: Comming soon

Key Contributions

  • Enhanced Pose Estimation and Filtering Pipeline
  • Biomechanically Grounded 2D Center of Mass (CoM) Calculation
  • Comprehensive Quantitative Validation
  • Demonstrated Practical Deployability

Datasets

  • Primary Dataset: 15 videos, 5,232 manually annotated frames from elite Canadian national team divers
  • Athletes: 4 elite divers (2 male, 2 female) from Canadian national team
  • Recording Location: Institut national du sport du Québec (INS Québec)
  • Annotations: 17-keypoint COCO poses with visibility labels, Center of Mass (CoM) ground truth computed using Winter's model, 8 joint angles (Left/Right elbow, shoulder, hip, knee), and kinematic events moments for 6 events per dive: 1) the highest point in the preparation phase, 2) touch down, 3) maximum springboard depression, 4) lift-off, 5) the highest point in the flight time, and 6) water contact moment
  • Dive Types: 8 distinct FINA dive codes (107B, 109C, 205B, 405B, 407C, 5152B, 5154B, 5337D) covering forward, backward, reverse, and inward groups with varying rotation counts and positions

Available Resources

All resources will be publicly released upon paper acceptance

Source Code

Complete Python implementation including:

  • Faster R-CNN with Feature Pyramid Network (FPN) for multi-object detection
  • HRNet-W48 for high-resolution 2D pose estimation
  • Two-stage temporal filtering pipeline (Hampel + Butterworth filters)
  • Winter's 2009 anthropometric CoM calculation with bilateral symmetry copying
  • Training scripts with data augmentation and cross-validation
  • Evaluation tools for PCK, CoM error, joint-angle error, and peak height metrics
GitHub Repository (Coming soon)

Trained Models

Pre-trained model checkpoints ready for deployment:

  • Faster R-CNN+FPN: Pre-trained detector for diver, springboard, and splash detection
  • HRNet-W48: Pre-trained pose estimator fine-tuned on diving dataset
  • 5-Fold Models: All models from cross-validation for reproducibility
  • Includes configuration files, filter parameters (Hampel: window=3, threshold=1.4826 Median Absolute Deviation (MAD); Butterworth: order=2, cutoff=15 Hz), and inference examples
Download Models (Coming soon)

Annotations & Datasets

Ground truth annotations in COCO JSON format:

  • 15 Videos: 5,232 frames with fully manual annotations
  • Bounding boxes for divers, springboards, and splash regions
  • 17-keypoint poses (COCO topology) with visibility labels for occluded joints
  • Center of mass ground truth computed from manual pose annotations using Winter's 2009 model
  • Joint angle ground truth for 8 joints (Left/Right elbow, shoulder, hip, knee)
  • Kinematic event labels for 6 key phases: preparation peak, touchdown, maximum springboard depression, liftoff, flight peak, water contact
  • Compatible with standard COCO dataset evaluation tools
Download Annotations (Coming soon)

Video Data

Raw diving video footage. The video data belongs to INS Québec and can be provided upon request subject to institutional approval:

  • 15 videos from Canadian national team training sessions at Institut national du sport du Québec (INS Québec)
  • 1920×1080 resolution, 100 FPS, captured with fixed high-speed camera positioned perpendicular to diving plane
  • 4 elite divers (2 male, 2 female), 8 unique FINA dive codes
  • Covers forward, backward, reverse, and inward dive groups with varying degrees of difficulty
  • Data Ownership: Videos belong to INS Québec
  • Access: Available upon request for research purposes with approval from INS Québec
Request Access (Contact after paper acceptance)

Experimental Results

Complete experimental data for reproducibility:

  • 5-fold cross-validation results with detailed metrics (PCK, CoM error, joint-angle error, peak height)
  • Per-joint PCK@0.25 breakdown for all 17 COCO keypoints
  • CoM error distributions (mean, median, std, 95th percentile) in pixels and centimeters
  • Joint-angle error statistics for all 8 joints with per-joint breakdowns
  • Comparison with DiveNet across all metrics
  • Evaluation scripts for generating all figures and tables in the paper
  • Raw metric CSVs for custom analysis
Download Results (Coming soon)

Publication Details

  • Title: Marker-less Estimation of Diver's Center of Mass Trajectory in Springboard Diving Using Deep Learning and a Single Fixed Camera
  • Authors: Ramon Figueiredo Pessoa, Rachid Aissaoui, Mathieu Charbonneau, Carlos Vazquez
  • Journal: Coming soon
  • Year: 2025
  • Funding: MITACS Accelerate program (grant IT07988)

Citation

BibTeX (will be updated upon publication):

@article{figueiredo2025divingfy,
  title={Marker-less Estimation of Diver's Center of Mass Trajectory in Springboard Diving Using Deep Learning and a Single Fixed Camera},
  author={Figueiredo Pessoa, Ramon and Aissaoui, Rachid and Charbonneau, Mathieu and Vazquez, Carlos},
  year={2025},
}

References

  • K. Murthy et al., "DiveNet: Dive Action Localization and Physical Pose Parameter Extraction for High Performance Training," IEEE Access, 2023.
  • D. A. Winter, Biomechanics and motor control of human movement. John Wiley & Sons, 2009, ch. 4: Anthropometry - Section: 4.1 Density, mass, and inertial properties.

Licensing

  • Code: MIT License (permissive open-source)
  • Annotations: Creative Commons Attribution 4.0 International (CC BY 4.0)
  • Pre-trained Models: Non-commercial research use only
  • Video Data: Owned by INS Québec, available upon request for approved research

Acknowledgments

This research was conducted at École de technologie supérieure (ÉTS), Université du Québec, in collaboration with the Institut national du sport du Québec (INS Québec). The study involved elite Canadian national diving team athletes and benefited from expertise provided by INS Québec coaches and biomechanists. The research received financial support from MITACS Accelerate program (grant IT07988).

Status Update

This page will be updated with active download links and detailed documentation once the paper receives acceptance notification. For any inquiries or requests for access to specific research collaborations, please contact Ramon Figueiredo via LinkedIn.

November 2025