Dataset Storage, Structured Artifacts, Split Schemas, and Model Serialization
Authoritative specification of file-based data structures and artifact persistence.
Note
SANKET does not utilize a traditional relational database (RDBMS) or NoSQL document store. As a real-time computer vision and machine learning system, its persistence layer is entirely file-based, structured artifact-driven, and version-controlled.
This architecture guarantees:
- Portability: Zero external database server dependencies; runs locally on any workstation or edge device.
- Reproducibility: Datasets, train/val/test splits, and evaluation manifests are represented as human-readable JSON/YAML files with explicit cryptographic checksums.
- Privacy Compliance: Raw video frames remain transient in memory; only abstract numerical coordinates and features are saved.
Sanket/
├── configs/
│ ├── default.yaml # Primary system configuration
│ └── splits/
│ └── signer_independent.json # Deterministic split definition
├── data/ # Local working data (Git-ignored)
│ ├── raw/ # Raw downloaded video archives (licensed)
│ ├── processed/ # Standardized frames / subclips
│ ├── landmarks/ # Extracted MediaPipe NumPy arrays (.npy)
│ └── features/ # Computed 645-dim feature matrices (.npz)
├── docs/
│ ├── datasets/
│ │ ├── DATASET_REGISTER.md # Authoritative dataset registry
│ │ └── INCLUDE_INSPECTION.md # Class balance and signer metadata
│ └── experiments/
│ └── manifests/ # Standardized evaluation JSON manifests
└── models/ # Serialized trained model estimators
├── baseline_random_forest.joblib
├── baseline_svm.joblib
└── baseline_mlp.joblib
Guarantees zero data leakage by strictly partitioning signers across sets:
{
"dataset_name": "INCLUDE_ISOLATED_16",
"version": "1.0.0",
"seed": 42,
"signer_split": {
"train_signers": ["signer_01", "signer_02", "signer_03", "signer_04", "signer_05"],
"val_signers": ["signer_06"],
"test_signers": ["signer_07"]
},
"sample_counts": {
"train": 480,
"val": 96,
"test": 96
},
"vocabulary_size": 16,
"sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
}Constructed by sanket.evaluation.manifest.EvaluationManifest:
{
"experiment_id": "EXP-P1-001-RF",
"timestamp": "2026-11-01T15:30:00Z",
"git_commit": "bd7aa10255c1b0a08295e407f5442bcc2b05e32e",
"model_type": "RandomForestSignClassifier",
"hyperparameters": {
"n_estimators": 100,
"max_depth": 15,
"random_state": 42
},
"dataset": {
"split_id": "signer_independent_v1",
"test_samples": 96,
"num_classes": 16
},
"metrics": {
"top_1_accuracy": 0.9333,
"top_5_accuracy": 1.0000,
"macro_precision": 0.9412,
"macro_recall": 0.9333,
"macro_f1": 0.9333,
"mcc": 0.9306
},
"latency": {
"median_ms": 0.56,
"p90_ms": 0.68,
"p95_ms": 0.73,
"p99_ms": 0.91,
"throughput_fps": 1507.2
}
}- Trained models are serialized using
joblib.dump(model, path, compress=3). - Compression level 3 balances disk size reduction with fast sub-second deserialization during runtime startup.
- Deserialization is protected by strict typing checks verifying that the unpickled object inherits from
BaseSignClassifier.