HRRP Missing Scenarios — Conditional Generation (DDPM / GAN)

Code release for our submission to EUSIPCO 2026 : Conditional Generative Models for High-Resolution Range Profiles: Capturing Geometry-Driven Trends in a Large-Scale Maritime Dataset

• Package: src/ship_hrrp_gen
• Configs: configs/
• Demo data: data/ship_hrrp.pt (128 HRRPs, length 512)

Figure 1 — Example HRRP samples from data/ship_hrrp.pt.

---

Overview

This repository focuses on conditional generative modeling for HRRPs. We study which acquisition and target parameters must be provided to the generator in order to:

1. generate ship-specific HRRPs (not just “HRRP-looking” signals), and
2. fill missing acquisition scenarios (especially missing aspect angles) while preserving physically meaningful trends.

The provided training entrypoint supports both GAN and DDPM backends (depending on the chosen YAML config).

---

Key contributions (what to look for)

1) Fundamental conditioning variables

We show that ship dimensions (length L, width W) and aspect angle (asp) are mandatory to generate ship-specific HRRPs. These conditions are interdependent (dimension effects cannot be properly modeled without angle, and vice versa), which is reflected in the performance table below.

Table 1 — Generation metrics for different models and conditioning types. Best scores per model are in bold.

2) Generated HRRPs follow a geometric projection law (TLOP)

Although HRRPs are noisy and difficult to interpret at the signal level, the apparent target extent along the range axis follows a simple geometric relationship with the aspect angle. We use the Theoretical Length of Object Projection (TLOP) model:

[ \mathrm{TLOP}(L, W, \mathrm{asp}) = |L\cos(\mathrm{asp})| + |W\sin(\mathrm{asp})| ]

where:

• L = true ship length (meters),
• W = true ship width (meters),
• asp = aspect angle at acquisition time (radians).

From an HRRP, we estimate the visual target length by detecting occupied range bins. This yields a measured quantity called Length on Range Profile (LRP). In real data, LRP correlates well with the TLOP curves.

Crucially, applying the same analysis to generated HRRPs shows that synthesized signals recover TLOP-consistent trends and can fill missing aspect-angle scenarios (at a coarse scale). This suggests the generator learns more than local realism: it preserves an underlying physical constraint linked to line-of-sight projection geometry.

Figure 2 — Correlation between measured/generated LRP and theoretical TLOP as a function of aspect angle.

---

Requirements

• Python ≥ 3.9
• PyTorch (CPU or CUDA)
• data/ship_hrrp.pt (demo dataset shipped with the repo)

---

Installation

cd github_repo
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install -e .   # installs the ship_hrrp_gen package from src/

Run a training

python -m ship_hrrp_gen.train \
  --config configs/gan_scalars_serloss.yaml \
  --data data/ship_hrrp.pt \
  --seed 42 \
  --num-workers 0

Useful flags:

• --num-workers: dataloader workers (set 0 on small CPUs).
• --skip-eval: skip test metrics if no test split.
• --fast-dev-run: quick pipeline sanity-check (1 train/val/test batch).

Quick sanity check

python -m ship_hrrp_gen.train \
  --config configs/gan_scalars_serloss.yaml \
  --data data/ship_hrrp.pt \
  --seed 0 \
  --num-workers 0 \
  --skip-eval \
  --fast-dev-run

Artifacts (checkpoints, figures, TensorBoard logs) are written under results/ following the figure_path in the config.

Layout

• src/ship_hrrp_gen/: models (DDPM, GAN), dataset, utils, training script (train.py).
• configs/: all YAML configs.
• requirements.txt: Python deps.
• .gitignore: ignores caches, venvs, and training outputs.

Notes

• Intended for a quick demo run on the 128 generated samples; no multi-GPU setup required.
• Final metrics rely on compute_metrics in ship_hrrp_gen.utils. If there is no test split (test_idx empty), use --skip-eval.

Citation

@inproceedings{brient2026eusipco,
  author    = {Edwyn Brient and Santiago Velasco{-}Forero and Rami Kassab},
  title     = {{Conditional Generative Models for High-Resolution Range Profile: Capturing Global Trends in a Large-Scale Maritime Dataset}},
  booktitle = {Proc. European Signal Processing Conference (EUSIPCO)},
  year      = {2026},
  note      = {to appear}
}