How pallavimitra/vision-based-absolute-pose-estimation-of-autonomous-vehicle works
Project type and purpose
• This is a vision-based machine-learning pipeline for 6-DoF pose estimation of cars from single images. It combines object detection (YOLOv3) with two regression stages: one for 3D coordinates (x,y,z) and one for yaw/pitch/roll.
Key components (from file tree and README)
• Data ingestion: raw images (images/) and Kaggle dataset.
• Preprocessing & utility functions: camera_utils.py, dataset.py, utils.py, batch_norm.py.
• Object detector: car_extractor.py wraps a pretrained YOLOv3 model (downloaded into trained_models/).
• Yaw/pitch/roll regressor: yaw_pred.py implements two-step angle estimation (classification into 8 bins + regression), applies camera matrix math from camera_utils.py.
• Orchestration notebooks:
• Car Position Estimation – Full pipeline.ipynb – runs detection, then coordinate and angle predictors end-to-end.
• Car-Yaw-Prediction.ipynb – trains or fine-tunes the yaw model.
• Models directory: trained_models/ (YOLO weights, pos_predictor.joblib, yaw model).
• Dependencies: Python, Jupyter, OpenCV, PyTorch/TensorFlow (for YOLO), scikit-learn, joblib, numpy.
High-level architectural pattern
• A linear, staged pipeline (sequence pattern), with clearly decoupled modules: detection → feature extraction → two parallel regressors → geometric projection.
Important architectural patterns & principles
• Modularity: separate detection, coordinate regression, angle regression.
• Reusability: utility scripts (camera_utils, utils) shared across stages.
• Data-flow clarity: linear pipeline with clear inputs/outputs between stages.