How thaonan/tracknetv5-sdk works
TrackNetV5 SDK supports tennis-ball detection and tracking, with V2/V4/WASB variants, Gaussian-heatmap data preparation, model training and video inference. The documented inference workflow processes video with a selected architecture and exports coordinates and visualizations; a separate core pipeline explicitly sequences detection, trajectory refinement and rendering. Training is configured through the factory-based runner. Edges are limited to README statements and shown calls or references; unsampled implementation details are not assumed.
Main entry points: builder.py Model Factory · train.py Training Entry · runner.py Training Runner · pipeline.py Tennis Pipeline
Components 31
Inference and Tracking
- Video Inference
track.py - Inference Module
infer.py - Heatmap Metrics
metrics_factory/metrics/tracknetv2_metric.py
Data Preparation
- Gaussian Preprocessing
tools/preprocess_data_gauss.py - Tennis Dataset
datasets_factory/datasets/tennis_dataset.py - Frame Transforms
datasets_factory/transforms/tracknet_transforms.py - Dataset Factory
datasets_factory/builder.py
Model Components
- Model Factory
models_factory/builder.py - TrackNet V2
models_factory/models/tracknetv2.py - TrackNet V4
models_factory/models/tracknetv4.py - TrackNet V5
models_factory/models/tracknetv5.py - WASB Model
models_factory/models/wasb.py - WASB HRNet
models_factory/backbones/wasb_hrnet_backbone.py - Motion Decoupling
models_factory/basic/mdd.py - Residual Refinement Head
models_factory/heads/r_strhead.py
Training and Evaluation
- Training Entry
train.py - Model Configurations
configs - Training Runner
engine/runner.py - Engine Factory
engine/builder.py - Training Loss
losses_factory/losses/tracknetv2_loss.py - Training Metric
metrics_factory/metrics/tracknetv2_metric.py - Optimizer Factory
optimizers_factory/builder.py - Visualization Hook
engine/hooks/visualizer_v2_hook.py
Detection Pipeline
- Tennis Pipeline
core/pipeline.py - Video Detector
core/detector.py - Trajectory Tracker
tracker/base.py - Video Visualizer
core/visualizer.py
Other components
- Video User
- Pipeline Caller
- Video Results
- Inference Exports
Connections 29
- Video User → Video Inference: runs on videos
README.md - Video Inference → Model Factory: builds architecture
track.py - Video Inference → Frame Transforms: uses transforms
track.py - Video Inference → Inference Exports: exports coordinates and video
README.md - Inference Module → Model Factory: builds model
infer.py - Inference Module → Frame Transforms: uses transforms
infer.py - Dataset Factory → Tennis Dataset: builds dataset
datasets_factory/datasets/tennis_dataset.py - Dataset Factory → Frame Transforms: builds transform pipeline
datasets_factory/transforms/tracknet_transforms.py - Model Factory → TrackNet V2: registers model
models_factory/models/tracknetv2.py - Model Factory → TrackNet V4: registers model
models_factory/models/tracknetv4.py - Model Factory → TrackNet V5: registers model
models_factory/models/tracknetv5.py - Model Factory → WASB Model: registers model
models_factory/models/wasb.py - Training Entry → Training Runner: starts training
train.py - Training Entry → Model Configurations: loads configurations
README.md - Training Entry → Dataset Factory: constructs data
train.py - Training Entry → Model Factory: constructs model
train.py - Training Entry → Optimizer Factory: constructs optimizer
train.py - Training Runner → Training Loss: computes loss
- Training Runner → Training Metric: evaluates predictions
- Training Runner → Engine Factory: configures hooks
- Training Runner → Visualization Hook: invokes hooks
- Training Metric → Heatmap Metrics: extracts centroids
engine/hooks/visualizer_v2_hook.py - Pipeline Caller → Tennis Pipeline: submits video
- Tennis Pipeline → Video Detector: detects points
core/pipeline.py - Tennis Pipeline → Trajectory Tracker: refines trajectory
core/pipeline.py - Tennis Pipeline → Video Visualizer: renders frames
core/pipeline.py - Tennis Pipeline → Video Results: saves refined video
core/pipeline.py - Video Detector → Model Factory: builds model
core/detector.py - Video Detector → Frame Transforms: uses transforms
core/detector.py