Aspect-Angle-Aware HRRP Classification

Reproducibility repository for the paper experiments on HRRP (High-Resolution Range Profile) classification with explicit conditioning on aspect angle.

Paper at a glance

This repository accompanies the paper "High-Resolution Range Profile Classifiers Require Aspect-Angle Awareness", submitted at EUSIPCO26.

The core message is about classification: HRRP classifiers perform significantly better when aspect angle is used as an explicit conditioning variable.

In our experiments, we compare unconditioned baselines vs. angle-aware variants (FiLM / CBN / concatenation) across multiple backbones. The paper reports a consistent gain in classification performance, with an average accuracy improvement of about 7% and gains up to 10% on some setups.

Overview

This repository is intended to reproduce the paper results:

• Conditional HRRP classification with aspect-angle-aware models
• Experiment configs used in the study (ResNet/CNN/MLP variants (single view), conditioning modes, GRU/LSTM/Transformer (multi-view)
• Train/validation/test evaluation workflow matching the reported setup
• Kalman-based aspect-angle notebook for secondary analysis and reproducibility checks

Quick Start

Installation

pip install -r requirements.txt
# Or, for development:
pip install -e .

Training

python src/clf_req_asp/train.py --config configs/arc_res/mstar_film.yaml --seed 42

Time-series HRRP (multi-profile sequence) training:

python src/clf_req_asp_ts/train.py --config configs_ts/ships_lstm_angle.yaml --seed 42

This command is provided for experiment reproduction (not as a general-purpose training framework).

Configuration

Training parameters are managed via YAML files in configs/ and configs_ts/:

• Dataset: MSTAR (10-class HRRP classification)
• Model: ResNet/CNN/MLP (single HRRP) or temporal GRU/LSTM/Transformer (HRRP sequences)
• Conditioning: Enable/disable aspect-angle conditioning
• Mechanism: FiLM, concatenation, or CBN

See configs/default.yaml for all available options.

Project Structure

.
├── src/clf_req_asp/             # Single-HRRP package (paper baselines + conditioned variants)
│   ├── models.py               # PyTorch model definitions
│   ├── train.py                # PyTorch Lightning training script
│   └── utils.py                # Data utilities and splitting
├── src/clf_req_asp_ts/          # Time-series HRRP package (multi-profile sequences)
│   ├── models.py
│   ├── dataset.py
│   ├── train.py
│   └── utils.py
│
├── configs/                     # Configuration files
│   ├── default.yaml            # Default config template
│   ├── arc_res/                # ResNetRP configs (MSTAR)
│   │   ├── mstar_film.yaml
│   │   ├── mstar_cbn.yaml
│   │   └── mstar_uncond.yaml
│   ├── arc_mlp/
│   │   ├── mstar_film_mlp.yaml
│   │   ├── mstar_cbn_mlp.yaml
│   │   └── mstar_uncond_mlp.yaml
│   └── arc_convs/
│       ├── mstar_film_cnn.yaml
│       ├── mstar_cbn_cnn.yaml
│       └── mstar_uncond_cnn.yaml
├── configs_ts/                  # Time-series configs (GRU/LSTM/Transformer)
│   ├── ships_lstm_angle.yaml
│   ├── ships_gru_angle.yaml
│   ├── ships_transformer_angle.yaml
│   ├── mstar_lstm_angle.yaml
│   ├── mstar_gru_angle.yaml
│   └── mstar_transformer_angle.yaml
│
├── aspect_estimation.ipynb      # Kalman-focused experiment notebook
│
├── README.md                    # This file
├── LICENSE                      # MIT License
├── requirements.txt             # Dependencies
└── pyproject.toml              # Package configuration

Models

All classifiers support optional aspect-angle conditioning.

Single-HRRP package (src/clf_req_asp):

• ResNetRP: Residual network with adaptive conditioning
• Conv1dBackbone: Lightweight 1D convolutional model
• MLPBackbone: Fully connected baseline

Time-series package (src/clf_req_asp_ts):

• ResNetRP + GRU temporal aggregation
• ResNetRP + LSTM temporal aggregation
• ResNetRP + Transformer temporal aggregation

Conditioning strategies:

• FiLM: Feature-wise Linear Modulation (scale + shift)
• cat: Concatenation of condition embedding with features
• CBN: Conditional Batch Normalization

Requirements

• Python 3.9+
• PyTorch 2.0+
• PyTorch Lightning 2.0+
• See requirements.txt for full list

Data

This work uses local data files under data/:

• data/MSTAR_hrrp.pkl for MSTAR experiments (dataset: MSTAR, key: data_path)
• data/ship_hrrp.pt for ship HRRP experiments (dataset: ships, key: path_rp)

The time-series package can directly consume data/ship_hrrp.pt via path_rp.

Example MSTAR config: configs/arc_res/mstar_film.yaml

Kalman notebook

aspect_estimation.ipynb is included for Kalman-specific experiments and reproducibility checks on aspect-angle estimation.
The main reproducibility runs are executed with src/clf_req_asp/train.py and the YAML files under configs/.

Training & Evaluation

Models are saved to results/{experiment_name}/checkpoints/best.ckpt.

Metrics logged to:

• TensorBoard: lightning_logs/{run_name}/
• CSV: results/Samples/{date}/metrics{hour}.csv

References

• Kalman Filtering: Constant-velocity state model with adaptive process noise
• Conditioning: FiLM (Perez et al., 2018), CBN (De Vries et al., 2017)

License

MIT License — see LICENSE for details.

Citation

If you use this code, please cite:

@misc{brient2026highresolutionrangeprofileclassifiers,
	title={High-Resolution Range Profile Classifiers Require Aspect-Angle Awareness}, 
	author={Edwyn Brient and Santiago Velasco-Forero and Rami Kassab},
	year={2026},
	eprint={2603.00087},
	archivePrefix={arXiv},
	primaryClass={eess.SP},
	url={https://arxiv.org/abs/2603.00087}, 
}