CRF-Net+ for INFRA-3DRC

CRF-Net+ for INFRA-3DRC adapts the CRF-Net camera-radar fusion framework for automotive object detection using the INFRA-3DRC dataset.

Main Contributions

• INFRA-3DRC dataset integration into CRF-Net
• Camera-radar fusion training across INFRA-3DRC 25 scenes
• INFRA-3DRC-specific data generator and configuration
• Automated training and evaluation reporting
• Per-class AP, precision, recall, and confusion matrices
• Prediction and ground-truth exports
• Training and evaluation visualisations

Dataset Configuration

Set the INFRA-3DRC dataset location in:

configs/infra3drc.cfg

For example:

[DATA]
data_set = infra3drc
data_path = /path/to/INFRA-3DRC

The calibration-corrected INFRA-3DRC dataset used is not included in this repository. The original INFRA-3DRC dataset is available from the official INFRA-3DRC dataset website

Training

python train_crfnet.py --config configs/infra3drc.cfg

Evaluation

Evaluation and automated reporting are implemented through:

evaluate_crfnet.py
utils/eval_test.py
utils/reporting.py

Generated outputs include performance metrics, per-class AP, confusion matrices, training curves, predictions, and ground-truth data.

Author

Olivier Rukundo, Ph.D.
University of Limerick, Ireland

Attribution

This repository is based on the original CRF-Net camera-radar fusion framework. Original CRF-Net authors/contributors retain authorship of the original implementation.

The INFRA-3DRC integration, dataset-specific modifications, and additional functionalities were developed by Dr Olivier Rukundo.

INFRA-3DRC is provided by Fraunhofer IVI.

License

This repository contains code derived from CRF-Net. Please refer to the included license and the original CRF-Net licensing terms.

Results

CRF-Net+ was trained and evaluated on all 25 INFRA-3DRC scenes.

mAP	Precision	Recall
0.9675	0.9956	0.5336

Evaluation Demo

The video below shows sequential object-detection outputs generated during evaluation.

CRFNetPlus_INFRA3DRC_Evaluation_H264_3.mp4