Initial commit add temporal fire tube experiments, an adapted impleme…

experiments/temporal-models/temporal-fire-tube/.dvc/.gitignore

@@ -0,0 +1,3 @@
+/config.local
+/tmp
+/cache

experiments/temporal-models/temporal-fire-tube/.dvc/config

@@ -0,0 +1,5 @@
+[core]
+    remote = s3remote
+    analytics = false
+['remote "s3remote"']
+    url = s3://pyro-vision-rd/dvc/experiments/temporal-fire-tube/

experiments/temporal-models/temporal-fire-tube/.dvcignore

@@ -0,0 +1,3 @@
+# Add patterns of files dvc should ignore, which could improve
+# the performance. Learn more at
+# https://dvc.org/doc/user-guide/dvcignore

experiments/temporal-models/temporal-fire-tube/.gitattributes

@@ -0,0 +1,2 @@
+*.ipynb filter=nbstripout
+*.ipynb diff=ipynb

experiments/temporal-models/temporal-fire-tube/.gitignore

@@ -0,0 +1,12 @@
+__pycache__/
+*.pyc
+.venv/
+.ipynb_checkpoints/
+*.egg-info/
+.pytest_cache/
+.ruff_cache/
+.cache/
+build/
+dist/
+htmlcov/
+.coverage

experiments/temporal-models/temporal-fire-tube/.python-version

@@ -0,0 +1 @@
+3.11

experiments/temporal-models/temporal-fire-tube/Makefile

@@ -0,0 +1,19 @@
+.PHONY: install lint format test help
+
+help: ## Show this help
+	@grep -E '^[a-zA-Z_-]+:.*?## .*$$' $(MAKEFILE_LIST) | sort | awk 'BEGIN {FS = ":.*?## "}; {printf "\033[36m%-15s\033[0m %s\n", $$1, $$2}'
+
+install: ## Install dependencies and set up notebook output stripping
+	uv sync
+	uv run nbstripout --install
+
+lint: ## Run ruff linter on code and notebooks
+	uv run ruff check .
+	@find notebooks/ -name '*.ipynb' 2>/dev/null | grep -q . && uv run nbqa ruff check notebooks/ || true
+
+format: ## Format code and notebooks with ruff
+	uv run ruff format .
+	@find notebooks/ -name '*.ipynb' 2>/dev/null | grep -q . && uv run nbqa ruff format notebooks/ || true
+
+test: ## Run tests with pytest
+	uv run pytest tests/ -v

experiments/temporal-models/temporal-fire-tube/README.md

@@ -0,0 +1,55 @@
+# Temporal Fire-Tube
+
+## Objective
+
+Reduce false positives from YOLO-based smoke/fire detection by applying a temporal
+verification step inspired by the fire-tube method (Park & Ko, 2020). This experiment builds 3D "tubes" from tracked detections across frames, extracts temporal features from each tube, and classifies with a Random Forest.
+
+## Approach
+
+Adapted two-step pipeline from "Two-Step Real-Time Night-Time Fire Detection
+Using Static ELASTIC-YOLOv3 and Temporal Fire-Tube" (Park & Ko, Sensors 2020):
+
+1. **YOLO inference** -- YOLOv11s detects smoke candidates in each frame.
+2. **Padding** -- Short sequences padded to minimum length by repeating boundary frames.
+3. **Tube construction** -- Detections tracked across frames via greedy IoU matching. For each tracked chain, the corresponding image regions are cropped and resized into a fire-tube.
+4. **Feature extraction** -- Tabular features (24 dims) computed per tube: area change, centroid shift, intensity change, histogram distance, confidence -- aggregated across consecutive pairs plus global features.
+5. **Classification** -- Random Forest (120 trees, depth 20) with balanced class weights classifies each tube as smoke/non-smoke.
+6. **Sequence decision** -- Positive if any tube is classified positive.
+
+**Key adaptation for Pyronear's 30s cadence:** The original paper uses optical flow on dense video (30fps). At 30-second intervals optical flow is meaningless, so we replace HoF features with tabular temporal features that capture how the detection region evolves over time.
+
+## Data
+
+Same dataset as `tracking-fsm-baseline` (shared `01_raw` DVC data):
+
+- Train: ~1,034 wildfire + ~1,433 false positive sequences (Pyronear only)
+- Val: ~112 wildfire + ~147 FP sequences (Pyronear only)
+- Ground truth: sequence-level binary labels
+
+## Results
+
+*TBD -- run `dvc repro` to generate.*
+
+## Pipeline
+
+```
+infer (01_raw -> 02_intermediate)
+  -> pad (02_intermediate -> 03_primary)
+    -> build_tubes (03_primary + 01_raw -> 04_feature)
+      -> extract_features (04_feature -> 05_model_input)
+        -> train (05_model_input/train -> 06_models)
+          -> predict (06_models + 05_model_input -> 07_model_output)
+            -> evaluate (07_model_output -> 08_reporting)
+```
+
+## How to Reproduce
+
+```bash
+make install          # Install dependencies
+dvc pull              # Download data from S3
+dvc repro             # Run full pipeline
+make lint             # Run linter
+make test             # Run tests
+```
+

experiments/temporal-models/temporal-fire-tube/dvc.yaml

@@ -0,0 +1,235 @@
+stages:
+  # --- YOLO Inference ---
+  infer:
+    foreach:
+      - train
+      - val
+    do:
+      desc: "Run YOLO inference on ${item} set images to produce raw smoke detections"
+      cmd: >-
+        uv run python scripts/infer.py
+        --data-dir data/01_raw/datasets/${item}
+        --model-path data/01_raw/models/${prepare.model_filename}
+        --output-dir data/02_intermediate/${item}
+        --confidence-threshold ${infer.confidence_threshold}
+        --iou-nms ${infer.iou_nms}
+        --image-size ${infer.image_size}
+      deps:
+        - scripts/infer.py
+        - src/data.py
+        - src/detector.py
+        - src/types.py
+        - data/01_raw/datasets/${item}
+        - data/01_raw/models/${prepare.model_filename}
+      params:
+        - prepare.model_filename
+        - infer
+      outs:
+        - data/02_intermediate/${item}
+
+  # --- Pad short sequences ---
+  pad:
+    foreach:
+      - train
+      - val
+    do:
+      desc: "Pad short ${item} sequences by repeating boundary frames"
+      cmd: >-
+        uv run python scripts/pad_sequences.py
+        --infer-dir data/02_intermediate/${item}
+        --output-dir data/03_primary/${item}
+        --min-sequence-length ${pad.min_sequence_length}
+      deps:
+        - scripts/pad_sequences.py
+        - src/data.py
+        - src/detector.py
+        - src/types.py
+        - data/02_intermediate/${item}
+      params:
+        - pad
+      outs:
+        - data/03_primary/${item}
+
+  # --- Build fire-tubes ---
+  build_tubes:
+    foreach:
+      - train
+      - val
+    do:
+      desc: "Build fire-tubes from ${item} set detections and images"
+      cmd: >-
+        uv run python scripts/build_tubes.py
+        --infer-dir data/03_primary/${item}
+        --data-dir data/01_raw/datasets/${item}
+        --output-dir data/04_feature/${item}
+        --crop-size ${tube.crop_size}
+        --max-tube-length ${tube.max_tube_length}
+        --confidence-threshold ${tube.confidence_threshold}
+        --max-detection-area ${tube.max_detection_area}
+        --iou-threshold ${tube.iou_threshold}
+      deps:
+        - scripts/build_tubes.py
+        - src/tube.py
+        - src/data.py
+        - src/detector.py
+        - src/types.py
+        - data/03_primary/${item}
+        - data/01_raw/datasets/${item}
+      params:
+        - tube
+      outs:
+        - data/04_feature/${item}
+
+  # --- Extract features ---
+  extract_features:
+    foreach:
+      - train
+      - val
+    do:
+      desc: "Extract tabular features from ${item} set fire-tubes"
+      cmd: >-
+        uv run python scripts/extract_features.py
+        --tube-dir data/04_feature/${item}
+        --output-dir data/05_model_input/${item}
+      deps:
+        - scripts/extract_features.py
+        - src/features.py
+        - src/types.py
+        - data/04_feature/${item}
+      params:
+        - features
+      outs:
+        - data/05_model_input/${item}
+
+  # --- Train classifier ---
+  train:
+    desc: "Train Random Forest classifier on tube features"
+    cmd: >-
+      uv run python scripts/train_classifier.py
+      --feature-dir data/05_model_input/train
+      --output-dir data/06_models
+      --n-estimators ${classifier.n_estimators}
+      --max-depth ${classifier.max_depth}
+      --random-state ${classifier.random_state}
+    deps:
+      - scripts/train_classifier.py
+      - src/classifier.py
+      - data/05_model_input/train
+    params:
+      - classifier
+    outs:
+      - data/06_models
+
+  # --- Predict ---
+  predict:
+    foreach:
+      - train
+      - val
+    do:
+      desc: "Run fire-tube predictions on ${item} set"
+      cmd: >-
+        uv run python scripts/predict.py
+        --feature-dir data/05_model_input/${item}
+        --model-dir data/06_models
+        --data-dir data/01_raw/datasets/${item}
+        --infer-dir data/03_primary/${item}
+        --output-dir data/07_model_output/${item}
+      deps:
+        - scripts/predict.py
+        - src/classifier.py
+        - src/data.py
+        - src/detector.py
+        - src/types.py
+        - data/05_model_input/${item}
+        - data/06_models
+        - data/01_raw/datasets/${item}
+        - data/03_primary/${item}
+      outs:
+        - data/07_model_output/${item}
+
+  # --- Package ---
+  package:
+    desc: "Bundle YOLO weights and classifier into a single deployable archive"
+    cmd: >-
+      uv run python scripts/package.py
+      --weights-path data/01_raw/models/${prepare.model_filename}
+      --classifier-path data/06_models/classifier.pkl
+      --params-path params.yaml
+      --output-path data/06_models/model.zip
+    deps:
+      - scripts/package.py
+      - src/package.py
+      - data/01_raw/models/${prepare.model_filename}
+      - data/06_models/classifier.pkl
+    params:
+      - prepare.model_filename
+      - infer
+      - tube
+    outs:
+      - data/06_models/model.zip
+
+  # --- Evaluate ---
+  evaluate:
+    foreach:
+      train_all:
+        split: train
+        subset: all
+        filter_flag: ""
+      val_all:
+        split: val
+        subset: all
+        filter_flag: ""
+      train_pyronear:
+        split: train
+        subset: pyronear
+        filter_flag: "--filter-prefix pyronear"
+      val_pyronear:
+        split: val
+        subset: pyronear
+        filter_flag: "--filter-prefix pyronear"
+    do:
+      desc: "Compute sequence-level metrics on ${item.split} set (${item.subset})"
+      cmd: >-
+        uv run python scripts/evaluate.py
+        --results-dir data/07_model_output/${item.split}
+        --output-dir data/08_reporting/${item.split}/${item.subset}
+        ${item.filter_flag}
+      deps:
+        - scripts/evaluate.py
+        - src/evaluator.py
+        - data/07_model_output/${item.split}
+      metrics:
+        - data/08_reporting/${item.split}/${item.subset}/metrics.json:
+            cache: false
+      plots:
+        - data/08_reporting/${item.split}/${item.subset}/plots
+
+  # --- Sweep ---
+  sweep:
+    foreach:
+      train_all:
+        split: train
+        subset: all
+        filter_flag: ""
+      train_pyronear:
+        split: train
+        subset: pyronear
+        filter_flag: "--filter-prefix pyronear"
+    do:
+      desc: "Grid-search fire-tube parameters on ${item.split} set (${item.subset})"
+      cmd: >-
+        uv run python scripts/sweep.py
+        --infer-dir data/03_primary/${item.split}
+        --data-dir data/01_raw/datasets/${item.split}
+        --output-dir data/08_reporting/sweep/${item.split}/${item.subset}
+        ${item.filter_flag}
+      deps:
+        - scripts/sweep.py
+        - src/tube.py
+        - src/features.py
+        - src/classifier.py
+        - src/types.py
+        - data/03_primary/${item.split}
+        - data/01_raw/datasets/${item.split}
+      outs:
+        - data/08_reporting/sweep/${item.split}/${item.subset}

experiments/temporal-models/temporal-fire-tube/params.yaml

@@ -0,0 +1,26 @@
+prepare:
+  model_repo: "pyronear/yolo11s_mighty-mongoose_v5.1.0"
+  model_filename: "yolo11s_mighty-mongoose_v5.1.0.pt"
+
+infer:
+  confidence_threshold: 0.01
+  iou_nms: 0.2
+  image_size: 1024
+
+pad:
+  min_sequence_length: 10
+
+tube:
+  crop_size: 64
+  max_tube_length: 50
+  confidence_threshold: 0.3
+  max_detection_area: 0.05
+  iou_threshold: 0.1
+
+features:
+  method: "tabular"
+
+classifier:
+  n_estimators: 120
+  max_depth: 20
+  random_state: 42