karung-counting-feedmill-semarang/docs/models.md

Models

Class names below are read directly from each checkpoint (YOLO(path).names). Formats present: .pt (PyTorch) → .onnx (ONNX) → .engine (TensorRT FP16 for Jetson; see export_model.py / export_v4.py).

Inventory

All weights live in models/ — full per-mode detector/filter matrix in models/modelREADME.md.

File	Classes	Task	Used by
models/truck-detector.{pt,engine}	{0: truck}	bbox	Dedicated truck specialist; src/config.py default truck model
models/model_karung_truk.{pt,onnx,engine}	{0: sack, 1: truck}	bbox	Legacy combined alias (auto-pick fallback in predict.py)
models/v4-best.{pt,onnx,engine} (v4-best (1).pt = duplicate copy)	{0: sack, 1: truck}	bbox	Combined model (Modes A/B/C/D truck; Modes A/C sack)
models/karung-dimuat-detection-di-feedmill-yolo26n-seg-200e.{pt,onnx,engine}	{0: person, 1: sack}	seg	Person-exclusion seg model (legacy predict_new.py)
models/yolo11n-bbox-100ep-sack+box-20260909-best.{pt,onnx,engine}	{0: sack, 1: box}	bbox	Unified sack+box (Modes B/C/D)
models/best.{pt,onnx,engine}	{0: sack}	seg	Sack-only specialist (Mode D sack; src/config.py default)

Model registry for the rpo_iki engine: rpo_iki/configs/model_registry.json (currently pins karung-dimuat-seg-200e, mAP50-mask 0.899, val MAE 0.67).

Support matrix

1. Multiple specialized models — YES (current design)

Two dedicated models run in parallel on the same frames:

• Truck model (truck-detector.pt) → presence, ROI, counting-line placement, batch lifecycle. Runs every 15th frame in src/main.py:130-145.
• Sack model (seg or combined) → per-frame track + count (src/main.py:163-169).
• predict_new.py:184-185 hardcodes the same duo (TRUCK_MODEL_PATH, SACK_MODEL_PATH); rpo_iki/count.py:1136-1143 prefers a dedicated truck-detector.pt for dynamic truck calibration.

2. Multiple classes within a single model — YES (filter-then-count)

• Combined models expose 2 classes (sack+truck, person+sack, sack+box).
• Consumers select what they need and ignore the rest:
  • SackDetector._parse (src/detection.py:34): keeps name == "sack" only.
  • ByteTrackTracker._parse (src/tracking.py:62): keeps ("sack", "truck").
  • predict_new.py:379-389,571-579: auto-detects karung/person class IDs by name substring, tracks classes=[person, sack], draws persons in red and skips them in counting.
  • rpo_iki/count.py (SACK_CLASS_ID = 1): counts one class; falls back to the same model's truck_class_id=2 when no dedicated truck model exists.
• Rule of thumb: only the sack/karung class is ever counted; other classes are auxiliary (ROI, visualization, exclusion).

3. Hybrid multi-model — YES (modes A/B/C/D, default C)

• predict.py --model-mode (or MODEL_MODE env / dashboard /api/batch/mode):

  Mode	Truck	Sack	Box
  A	v4 (sack+truck)	v4	—
  B	v4 (truck-only)	yolo11n	yolo11n
  C (default)	v4 (sack+truck)	v4	yolo11n (box-only)

• All weights load as .engine (verified coexist: 2 engines ~16 MB, 3 engines ~24 MB peak of 7.6 GB). Dashboard switches persist to batch_mode.json and apply on next service restart (models load once at startup).

• Counting: MultiClassLineCounter = dual LineCrossCounters on one shared line; events tagged with class_name; sack and box track IDs live in separate spaces.

Choosing / swapping a model

• src/ pipeline: MODEL_SACK_PATH / MODEL_TRUCK_PATH env vars (defaults: ./models/best.engine / ./models/truck-detector.engine; see configuration.md) or --source for files.
• predict.py: config.yaml models.paths is canonical (per-mode presets in models.modes, per-class conf/iou/min_bbox in models.detection_params). Legacy overrides still work: MODEL_PATH env or --model (single v4 file), --box-model (yolo11n weights), --model-mode / deprecated MODEL_MODE env for the preset, --sack-conf / --truck-conf / --box-conf for thresholds.
• TensorRT: python export_model.py models/<name>.pt (FP16 .engine) on the Jetson; production loads .engine only (missing .engine falls back to the .pt sibling with a warning). deploy_to_jetson.py syncs config.yaml + the .engine files.