Skip to content

MFSR Project Wiki — Index

Articles (Concepts & Research)

  • PIUnet Architecture — Technical deep-dive on PIUnet's TEFA attention blocks, TERN alignment module, residual learning variants, and LR fusion variants; includes comparison to modern approaches and identified weaknesses
  • Existing Data Limitations — Problems with flight 21051/21052 data: sparse overlap, single site, training/inference gap, parallax, non-designed SR factor. The case for new data collection.
  • Data Collection v2 — Proposed new data collection: 2x magnification math, flight parameters, site selection criteria, Metashape integration.
  • Evaluation Strategy — Revised evaluation approach: LR-consistency check, sharpness gain, why PSNR and LPIPS are problematic for LWIR, self-consistency checks.
  • Autoresearch for SR — Adapting Karpathy's autoresearch for LWIR SR: 3-file architecture, 10-min budget, PSNR metric, anti-cheating, phased approach.
  • HighRes-net Paper — The original HighRes-net paper (ICLR 2020): recursive fusion, implicit co-registration via median reference, ShiftNet registration-at-the-loss, ESA Kelvin competition context, and why its satellite-centric assumptions broke down for LWIR aerial imagery.
  • Project Timeline — Chronological narrative of the entire LWIR MFSR project from May 2018 to present, reconstructed from git histories across all 11 repositories. Covers the progression from classical SR to HighRes-net to PIUnet, the registration bottleneck, tool proliferation, and the bicubic gap.

Projects (Tools & Components)

  • PIUnet — PIUnet MFSR neural network for LWIR thermal super-resolution; five model variants (original, residual V1/V2, LR fusion V1/V2), could not beat bicubic baseline, project paused Nov 2025
  • Hephaestus — PyQt6+OpenGL interactive thermal image registration GUI; LWIR-to-LWIR (Phase 1 complete) and LWIR-to-VNIR cross-modality alignment (Phase 2 in progress)
  • LWIR-Align — Registration pipeline for 675 HA thermal frames to LA mosaics; SuperGlue + RAFT methods, group-based processing, tile metadata generation
  • LWIR Tile Validator — PyQt6 GUI for systematic validation of HA-to-mosaic tile alignment; GPU-rendered patch grid, annotation system, ProbaV dataset export
  • HighRes-net — HighRes-net (ESA Kelvin winner) fork, first MFSR architecture tried for LWIR; extensive custom architecture work (EnhancedProgressiveMFSRNet), 0 dB improvement over bicubic on LWIR, abandoned Apr 2025 in favor of PIUnet
  • MFSR Mosaic — Georeferencing, tiling, and HA/LA patch pair extraction pipeline; INS calibration, KML-to-GeoTIFF conversion, SuperGlue alignment planned
  • ProbaV Data — ProbaV satellite super-resolution dataset: the origin point of the MFSR project, its processed state with enhanced SuperGlue registration, and how its directory format became the canonical data interchange standard for all LWIR tooling
  • Aerofocus — PyQt6 blur detection and focus quality assessment tool for LWIR frames; Laplacian variance and FFT analysis, motion data correlation, built April 2025 in 4-day sprint, prototype complete but frame rejection pipeline not integrated

Struggle Logs

  • Training Instability — End-to-end PIUnet training produced wildly inconsistent results due to gradient competition between alignment and reconstruction modules, compounded by poor experiment tracking.
  • Bicubic Gap — PIUnet scored 0.35-0.99 dB below bicubic on all 3 test sequences at the Nov 2025 boss meeting, caused by spatially-invariant alignment (TERN), registration-sensitive metrics, and limited training data.
  • Inference Evaluation — Full-frame inference evaluation was unreliable because every step from SR output to mosaic comparison added registration noise, and only 3 test sequences had sufficient frame overlap.
  • Experiment Tracking — W&B logged metrics but not code versions, full configs, or dataset versions, making it impossible to diagnose why training runs varied or reproduce successful results.
  • PQ Encoding (Not Applicable) — Why PQ/PU21 encoding does NOT help narrow-range LWIR data despite headline 2-9 dB claims on HDR imagery.