Klip Postprocessing Eval

klip-postprocessing-eval

A PyTorch Benchmark for High-Contrast Imaging Post Processing — Ko, Douglas, and Hom (2024) (arXiv:2409.16466, 2024)

What this evaluates

Evaluates the accuracy and computational efficiency of PCA-based PSF subtraction algorithms for high-contrast astronomical imaging. Specifically, it measures how well the algorithm mitigates speckle noise while recovering planetary signals compared to a reference implementation.

Datasets

• Beta Pictoris ($\beta$ Pic) — total ?; splits: test (-1)
• HR8799 — total ?; splits: test (-1)

Metrics

• SNR (primary) — range: ratio
  • Signal-to-Noise Ratio calculated for each detected planet in the processed image. Defined as the peak planet flux divided by the local background noise standard deviation.
• computation time — range: seconds
  • Wall-clock time required to run full-frame PCA PSF subtraction on the dataset.

Input / output format

Input: Multi-frame astronomical image cubes (sequences) for the Beta Pictoris and HR8799 systems.

Output: Processed single-frame images after PSF subtraction, SNR values for detected planets, and total processing time.

Scoring recipe

def evaluate(dataset, model_output):
    # model_output contains processed image and metadata
    planets = detect_planets(model_output.image)
    snr_values = []
    for planet in planets:
        snr = planet.peak_flux / local_background_noise_std(model_output.image, planet.mask)
        snr_values.append(snr)
    time_taken = model_output.computation_time
    return {"snr": snr_values, "time": time_taken}

Common pitfalls

• The benchmark only compares full-frame PCA, ignoring pyKLIP's additional annular and subsection features that could improve SNR but increase runtime.
• SNR values are measured only for successfully detected planets; masked gray regions in the images are excluded from the calculation.

Evidence (verbatim from paper)

To benchmark our package torchKLIP against pyKLIP, we compared the SNR and computation time using full-frame PCA on the $\beta$ Pic and HR8799 datasets.

Citation

@misc{ko2024torchklip,
  title={A PyTorch Benchmark for High-Contrast Imaging Post Processing},
  author={Ko, Douglas, and Hom (2024)},
  year={2024},
  note={arXiv:2409.16466}
}

• arXiv: 2409.16466