Ready-made dataset

Professional AI Motion & Pose Dataset - 25k+ HQ Frames

The Ultimate High-Density Locomotion Datasetfor Computer Vision & AI Training.

This is a professional-grade dataset engineered for Human Pose Estimation (HPE) and Motion Analysis. It features 106 diverse video sources processed with a rigorous quality-first pipeline.

One-time: $350Source videos: 106Total rows: 25,944 (incl. augments)Accepted (pre-aug): 8,648

Dataset Overview

  • Source: 106 high-quality source videos (focused on running/locomotion)
  • Quantity: 25,944 total frames (including augmentations)
  • Original Frames: 8,648 strictly filtered HQ frames (pre-augmentation)
  • Quality Standard: Mean quality score of 0.8578/1.0
  • Acceptance Rate: 58.40% (only high-visibility frames exported)

Multi-Layer Technical Architecture

  • Layer 1.0 (Pose): 33-landmark MediaPipe/BlazePose set (x, y, z, visibility)
  • Layer 1.1: RGB mean/std, brightness, contrast, edge density, and 12-bin hue histograms
  • Temporal Smoothing: Gaussian filter (window=5) for jitter-free sequence analysis
  • Body Normalization: Hip-centered + torso-scale normalization fields included
  • Kinematic Features: Per-frame velocity, acceleration, and stride length proxies

Structured data layout

  • data.jsonl — global merged dataset for training
  • per_video/*.jsonl — per-source splits for sequence testing
  • features.json — sequence-level metrics (motion consistency, step frequency)
  • low_quality_frames.jsonl — rejected frames for hard-example mining / negative training
  • manifest.json, global_stats.json, export_quality_report.json — metadata and QA

Free samples

Validate the format on GitHub: QualityVision-Motion-Dataset-Samples.

Purchase & licensing

  • Checkout: Gumroad (one-time $350; see product page)
  • Deliverable: JSONL + metadata exports — not raw source videos
  • License: Proprietary / commercial (single-seat; see Gumroad)
  • Contact: info@qvision.space