Louis Forster

Back to portfolio

Low-Light Video Denoising

Low-light video denoising research. Currently under review at IEEE ICME as second author. Beating previous SOTA with 29.22 dB PSNR and 0.863 SSIM on BVI-RLV with bidirectional warping and adaptive occlusion masking.

Optical flow fields from DPFlow (base), DPFlow-Noisy (fine-tuned, middle), and RAFT. Fine-tuning cut temporal residual MSE from 2.94 to 2.75 — sharper edges, less bleed into occluded regions.
Optical flow fields from DPFlow (base), DPFlow-Noisy (fine-tuned, middle), and RAFT. Fine-tuning cut temporal residual MSE from 2.94 to 2.75 — sharper edges, less bleed into occluded regions.

Impact

  • Under review at IEEE ICME as second author
  • Reached 29.22 dB PSNR, 0.863 SSIM, and 0.334 LPIPS on BVI-RLV, all above the pre-tuned baseline
  • Fine-tuned model cut temporal residual MSE from 2.9417 (DPFlow base) to 2.7472 — a 6.6% reduction

Stack

PythonPyTorchPTLFlowRAFTOpenCVLPIPS

Key Decisions

  • Occlusion handling: Confidence-weighted fusion of forward and backward warps reduced artifacts in ambiguous regions.
  • Robustness: Explicit fallback paths kept long sequence runs alive when bidirectional warping failed on edge cases.
  • Evaluation discipline: Sequence-level and averaged metrics made model regressions immediately visible.

Technical Approach

  • Each frame is denoised/enhanced while temporal hidden states are cached across a sequence.
  • Optical flow is computed in both directions using RAFT/PTLFlow to align temporal context.
  • Occlusion masks are derived from forward-backward consistency error with an adaptive threshold.
  • Forward/backward warped candidates are fused using photometric confidence weighting.
  • A pipeline runner orchestrates training and evaluation and exports per-sequence plus averaged metrics.

Visual Comparison

Model comparison visual

Project Links