How nampham13/bt-shutternet works
BT-ShutterNet estimates exposure fraction (and shutter angle) from blurred video by comparing inter-frame motion with within-exposure blur. The principal workflow runs from a clip or supervised frame/flow data, through preprocessing and dataset selection, to model inference and reported estimates or evaluation metrics. The model’s useful internal layers are its temporal-motion branch, key-frame blur-physics branch, and weighted physics estimator. The graph also preserves documented training, benchmarking, and data-visualization capabilities. Relationships are limited to README statements, shown imports, or SOURCE INDEX references; unsampled internals are not treated as inspected.
Main entry points: datasets/factory.py Dataset selection · models/factory.py Model configuration · preprocessing.py ROI preprocessing · bt_net.py Key-frame blur
Components 21
Data preparation
- ROI preprocessing
utils/preprocessing.py - ROI operations
utils/roi.py - Synthetic blur dataset
datasets/exposure_flow_dataset.py - Real video dataset
datasets/exposure_video_dataset.py - Dataset selection
datasets/factory.py
Exposure model
- Model configuration
models/factory.py - Encoder and blocks
models/blocks.py - Blur descriptor
models/blur_descriptor.py - Temporal motion
models/bt_net.py - Key-frame blur
models/bt_net.py - Physics estimator
models/bt_net.py
Train and evaluate
- Training
scripts/train.py - Evaluation
scripts/evaluate.py - Video inference
scripts/infer_video.py - Exposure metrics
utils/metrics.py
Diagnostics
- Model benchmark
scripts/benchmark.py - Data visualization
utils/visualization.py - Visualization workflow
scripts/visualize_data.py
Other components
- Video user
- Frame and flow data
- Video manifest
Connections 25
- Video user → Video inference: submits clip
README.md - Frame and flow data → Synthetic blur dataset: supplies frames and flow
README.md - Video manifest → Real video dataset: lists video samples
datasets/exposure_video_dataset.py - Synthetic blur dataset → ROI preprocessing: preprocesses frames
datasets/exposure_flow_dataset.py - Real video dataset → ROI preprocessing: preprocesses frames
datasets/exposure_video_dataset.py - ROI preprocessing → ROI operations: uses ROI operations
utils/preprocessing.py - Dataset selection → Synthetic blur dataset: selects dataset
datasets/factory.py - Dataset selection → Real video dataset: selects dataset
datasets/factory.py - Training → Dataset selection: creates dataset
scripts/train.py - Evaluation → Dataset selection: creates dataset
scripts/evaluate.py - Model configuration → Temporal motion: constructs model
models/factory.py - Temporal motion → Encoder and blocks: uses blocks
models/bt_net.py - Key-frame blur → Blur descriptor: uses descriptor
models/bt_net.py - Key-frame blur → Encoder and blocks: uses blocks
models/bt_net.py - Temporal motion → Physics estimator: provides motion cues
README.md - Key-frame blur → Physics estimator: provides blur cues
README.md - Training → Model configuration: builds model
scripts/train.py - Evaluation → Model configuration: builds model
scripts/evaluate.py - Video inference → Model configuration: builds model
scripts/infer_video.py - Training → Exposure metrics: reports metrics
scripts/train.py - Evaluation → Exposure metrics: computes metrics
scripts/evaluate.py - Video inference → ROI preprocessing: preprocesses clip
scripts/infer_video.py - Model benchmark → Model configuration: benchmarks model
- Visualization workflow → Dataset selection: creates dataset
scripts/visualize_data.py - Visualization workflow → Data visualization: renders samples
scripts/visualize_data.py