Diagram Alur Arsitektur & Pipeline Data Sistem Retraining Karung Diagram Alur Arsitektur 8-Tahap: Video Ingest, FFmpeg Slicing, SAM3 Grounding, Review Canvas, Data Prep & Triage, Master Dataset, YOLO Retraining, Live Inference & Benchmark 2026-08-27T00:00:00Z Retraining Studio FODG Specification Generator Sistem Retraining & Live Counter Feedmill — Arsitektur & Pipeline Data 8 Tahap Pipeline Closed-Loop End-to-End: Ingest CCTV 06:00, Ekstraksi FFmpeg, Auto-Label SAM3, Review Canvas, Triage Data, Dataset Freezing, YOLO Retraining & Live Inference 1. Video Archive & CCTV Ingest Siklus 24 Jam (06:00 Cutoff) & Ingestion [RTSP :8554] [06:00 Cutoff] [Sidecar JSON] • Ingest stream Dahua RTSP :8554 / WHEP :8889 • Pengenalan OCR timestamp video otomatis • Partisi batch per hari kerja feedmill (06:00-05:59) • Metadata sidecar JSON & storage read-only IN: RTSP Stream / Batch MP4 | OUT: Indexed Video 2. Video Library & FFmpeg Slicing Sub-Sampling Engine & Frame Extraction [FFmpeg 1.0 FPS] [Range Trim] [%06d.jpg] • Range streaming & visual time-range trimmer • Ekstraksi frame presisi (default: 1.0 FPS) • Penulisan sekuens zero-padded 000001.jpg • Atomic write ke data/batches/<batch-id>/frames/ IN: Video Batch + Range | OUT: JPG Frames 3. SAM3 Grounding Engine Singleton CUDA & Zero-Shot Auto-Label [Singleton GPU] [1 set_image] [Exemplar Pool] • Singleton CUDA manager (1 set_image per frame) • Multi-prompt zero-shot grounding per kelas • Cross-prompt NMS & ekstraksi polygon mask • Exemplar visual tuning (positive/negative points) IN: JPG Frames + Prompts | OUT: Polygon Labels 4. Review Canvas & Editor Dark-Mode Interactive Workspace [Dark Canvas] [Vertex/BBox] [Hotkeys Space/W/D] • Manipulasi interaktif vertex polygon & bounding box • Hotkey efisiensi (Space approve, W triage, D delete) • Re-assignment kelas label & filmstrip quick reclass • Pelacakan sumber (auto SAM3 vs manual edit) IN: Candidate Labels | OUT: Approved Annotations 5. Data Prep & Outlier Triage Multi-Signal Quality Filter & Augment [Scatter Plot] [Crop Grid] [Augment Presets] • Filter keep-ranges (score, area_pct, aspect_ratio) • Scatter plot interaktif (Score x Area) & crop grid • Manual override per bounding box (keep/ignore) • Preset augmentasi (Off, Light, Medium, Aggressive) IN: Approved Batches | OUT: Filtered Annotations 6. Master Dataset Freezing Deterministic Immutable Repository [SHA1 Stable Val] [Immutable] [YOLO TXT] • Val split stabil permanen berbasis hash SHA1 • Ekspor label format YOLO TXT (class x y w h) • Generasi manifest data.yaml & snapshot rules • Integrasi base datasets eksternal (train-only) IN: Filtered Labels | OUT: dataset/ train/ val/ 7. YOLO Retraining & Benchmark Ultralytics Fine-Tuning & Evaluation [Ultralytics YOLO11] [mAP Compare] [Auto VRAM] • Fine-tuning YOLO11 dari base model terdaftar • VRAM auto-tuning & hardware agnostic execution • Evaluasi otomatis metrik mAP50 & mAP50-95 • Ekspor model terbaik best.pt & log metrics.json IN: Master Dataset | OUT: best.pt + Metrics 8. Live Inference & Counter Real-Time Belt Counting & Benchmark [ByteTrack] [LineCross y1] [WHEP Stream :8889] • ByteTrack tracking & tripwire line cross (y1 edge) • Deploy best.pt ke runtime menghitung-karung • Stream low-latency WHEP di UI operator (:5000) • Headless benchmark otomatis vs Ground Truth Excel IN: best.pt + Live Video | OUT: Real-Time Counts Infrastruktur & Storage Bus Terintegrasi • Storage Hierarchy: data/app.db (SQLite WAL) | data/projects/<slug>/dataset/ (Images, Labels, data.yaml) | data/batches/ (Frames 1.0 FPS) | models/<v>/best.pt • Port Matrix: Web UI (:8080) | FastAPI REST API (:8000) | Frontend Dev (:5173) | Operator Live Counter (:5000) | CCTV RTSP (:8554) | MediaMTX WHEP (:8889) • Operational Invariants: 1. Val Split Stabil Permanen (SHA1 Hash) | 2. 1 set_image() Per Image di SAM3 | 3. Video Archive Read-Only | 4. Dynamic GPU Auto-Tuning