Bridging the Gap Between Computational Photography and Visual Recognition: 7th UG2+ Workshop & Challenge

Event: CVPR 2024 Workshop · Duration: 136 min · ▶ Watch on YouTube

Abstract

The 7th UG2+ Workshop & Challenge focuses on bridging the gap between computational photography and visual recognition. The workshop features invited talks from leading researchers in computational imaging and computer vision, covering topics such as health equity, low-light imaging, and face recognition in low-quality imagery. It also presents results from three challenge tracks: atmospheric turbulence mitigation, all-weather semantic segmentation, and UAV tracking/pose-estimation. A key theme is the development of robust and trustworthy algorithms that can perform well under non-ideal visual conditions and provide reliable uncertainty quantification for scientific applications.

Speakers

  • Achuta Kadambi — Associate Prof. @ UCLA
  • Kristina Monakhova — Postdoctoral Fellow, MIT (Incoming Assistant Professor @ Cornell CS)
  • Minchul Kim — Ph.D. student @ MSU

Talks (3)

  • 00:25:59 — Achuta Kadambi: On Health Equity and Computational Imaging
    • This talk discusses how computational imaging can address issues algorithms alone cannot solve, using blood volume estimation as a running example, and explores biases in light-based sensors and solutions using radar.
  • 00:33:02 — Kristina Monakhova: Learning noise models and error bars for low light imaging
    • This talk presents methods for learning noise models for extremely low-light videography and quantifying uncertainty in denoising for multiphoton microscopy, emphasizing the importance of trustworthy algorithms for scientific imaging.
  • 01:32:05 — Minchul Kim: Face Recognition in Low-Quality Imagery
    • This talk explores methods for robust face recognition in low-quality imagery, focusing on adaptive margin loss functions, feature fusion techniques for video, and keypoint relative position encoding to improve performance and robustness.

Key Takeaways

  • Computational imaging can address biases in light-based sensors and improve health equity by combining different modalities like radar and light.
  • Developing physics-informed GANs for noise modeling allows for the creation of realistic synthetic noisy data, crucial for training robust denoisers for extreme low-light conditions.
  • Quantifying uncertainty in deep learning reconstructions is vital for scientific imaging, enabling the detection of model hallucinations and driving adaptive acquisition strategies in microscopy.
  • Face recognition models can be made more robust to low-quality images by using adaptive margin loss functions that emphasize hard but identifiable samples and de-emphasize unidentifiable ones.
  • Feature fusion techniques, particularly two-stage paradigms involving clustering and aggregation, are effective for handling variable-length video inputs in face recognition, improving performance and efficiency.

Methods / Models / Datasets Mentioned

  • AdaFace
  • CAFace
  • KP-RPE
  • V-BM4D
  • FastDVDnet
  • L2SID
  • Physics-based noise GAN
  • Quantile regression
  • Conformal risk control
  • Bayesian Dropout
  • Ensemble techniques
  • Blood Decomposition

Topics

Computational Photography · Visual Recognition · Atmospheric Turbulence Mitigation · Semantic Segmentation · UAV Tracking and Pose Estimation · Low Light Imaging · Noise Modeling · Uncertainty Quantification · Face Recognition · Feature Fusion

Notes

Open for commentary — connections to other work, critiques, follow-up reading.