Dataset Pipeline Overview

This project uses a 4-step pipeline driven by a shared config.json:

1. Generate (Generate.py): Renders synthetic images and COCO annotations to paths.output_base_dir.
2. Annotate (Annotate.py): Converts COCO to OBB labels, creates multiple edge/style views, and prepares datasets in white/black splits.
3. Train (train.py): Trains YOLO-OBB models for selected styles using dataset paths and training settings.
4. Benchmark (Benchmark_Metrics.py): Evaluates trained models on test sets and exports per-class metrics and LaTeX tables.

Only the parameters listed below are read from config.json. All other options remain in-script.

Generate.py — Config parameters used

• paths.scene_blend_file: Absolute path to the Blender scene .blend file.
• paths.category_map_file: Path to the category map JSON for annotation names.
• paths.output_base_dir: Base directory where generated COCO dataset is written.
• model.color: Hex color (e.g., #0f0f13) applied to the imported model.
• model.model_path: Absolute path to the .stl model to import.
• timing.start_time: Optional epoch start time used for progress reporting. If null, uses current time.
• timing.initial_count: Baseline image count for generation progress.

Run: blenderproc run /home/reddy/Bachelor_Thesis/Generate.py

Annotate.py — Config parameters used

• paths.output_base_dir: Base dir to read generated COCO data from.
• paths.dataset_white_dir: Destination root for "white" dataset variant.
• paths.dataset_black_dir: Destination root for "black" dataset variant.

Run: python /home/reddy/Bachelor_Thesis/Annotate.py

train.py — Config parameters used

• training.model: Style to train (e.g., control, canny, ...).
• training.dataset_path: Dataset root used for training.
• training.model_size: YOLO size key (e.g., n, s, m).
• training.epochs: Number of epochs.
• training.imgsz: Image size.
• training.patience: Early stopping patience.
• training.batch: Batch size.
• training.project_suffix: Suffix segment for the output training directory name.
• training.yolo_config_pattern: Pattern for model cfg, e.g., yolo11{size}-obb.yaml.
• training.yolo_weights_pattern: Pattern for pretrained weights, e.g., yolo11{size}.pt.

Run: python /home/reddy/Bachelor_Thesis/train.py

Benchmark_Metrics.py — Config parameters used

• paths.test_sets_dir: Root folder containing test sets (with images/, labels/, data.yaml).
• paths.trains_base_dir: Base directory of trained model runs (used to locate weights/best.pt).
• paths.benchmarks_base_dir: Base directory to write benchmark outputs and LaTeX.

Run: python /home/reddy/Bachelor_Thesis/Benchmark_Metrics.py