How cknishan/synthetic-squat-form-checker works
This repository is a synthetic squat-form classification research pipeline, not a dependable exercise-feedback product. The documented workflow generates controlled Blender clips, extracts and validates MediaPipe poses, derives features and grouped splits, then compares rule-based, logistic, clip-MLP and temporal-CNN predictions. It also supports frozen-model exploratory inference on real videos. The README documents the broad pipeline; sampled excerpts and the source index support specific internal links. Uninspected modules and generated-data wiring are represented only where documented.
Main entry points: create_milestone_splits.py Grouped splits · compare_real_models.py Real-video comparison · render_milestone_dataset.py Controlled rendering · build_milestone_features.py Clip and frame features
Components 19
Synthetic generation
- Character preparation
scripts/blender/prepare_milestone_characters.py - Motion variants
scripts/blender/create_milestone_variants.py - Controlled rendering
scripts/blender/render_milestone_dataset.py
Pose and features
- Pose extraction
scripts/data/extract_milestone_landmarks.py - Pose quality checks
scripts/data/validate_milestone_landmarks.py - Render validation
scripts/data/validate_rendered_dataset.py - Clip and frame features
scripts/data/build_milestone_features.py - Grouped splits
scripts/data/create_milestone_splits.py - Pilot feature plots
scripts/data/plot_pilot_features.py
Model evaluation
- Rules and logistic
scripts/models/train_baselines.py - Clip MLP
scripts/models/train_mlp.py - Temporal CNN
scripts/models/train_temporal.py
Inference
- Video sampling
scripts/data/inference_video.py - Video pose extraction
scripts/data/extract_mediapipe_landmarks.py - Inference features
scripts/data/inference_features.py - Logistic prediction
scripts/models/primary_logistic_inference.py - Real-video comparison
scripts/data/compare_real_models.py
Other components
- Researcher
- Real-video investigator
Connections 19
- Researcher → Character preparation: prepares scene
README.md - Character preparation → Motion variants: provides scene
README.md - Motion variants → Controlled rendering: provides actions
README.md - Controlled rendering → Pose extraction: renders clips
README.md - Pose extraction → Pose quality checks: validates poses
scripts/data/extract_milestone_landmarks.py - Controlled rendering → Render validation: validates renders
scripts/blender/render_milestone_dataset.py - Pose quality checks → Clip and frame features: checks landmarks
scripts/data/build_milestone_features.py - Clip and frame features → Grouped splits: supplies features
scripts/data/create_milestone_splits.py - Clip and frame features → Pilot feature plots: plots features
scripts/data/build_milestone_features.py - Grouped splits → Rules and logistic: partitions data
scripts/models/train_baselines.py - Grouped splits → Clip MLP: partitions data
scripts/models/train_mlp.py - Grouped splits → Temporal CNN: partitions data
scripts/models/train_temporal.py - Real-video investigator → Video sampling: supplies video
README.md - Video sampling → Video pose extraction: samples frames
scripts/data/inference_video.py - Video pose extraction → Inference features: provides landmarks
- Inference features → Logistic prediction: provides predictors
scripts/data/inference_features.py - Real-video comparison → Logistic prediction: loads model
scripts/data/compare_real_models.py - Real-video comparison → Inference features: uses features
scripts/data/compare_real_models.py - Real-video comparison → Temporal CNN: compares models
scripts/data/compare_real_models.py