[johnaoga]

Summary

Addresses two issues found while running the apps:

1. singledic was ~18× slower than xdic (≈280 s/frame vs ≈15 s/frame) even though it tracks a single camera.
2. proxyncorr couldn't consume xDIC-style .mat ROI/seed sidecars and exposed only a subset of the engine's DIC modes.

1. singledic slowness — fixed

Root cause: SingleDicWorkflow::track() drove the engine through ncorr::NcorrSession::process_frame() one frame at a time. Each call ran plain DIC_analysis({ref, def}), which:

• re-ran the full automatic seed search from scratch on every frame — the seed computed via loadOrCreateSeed() was never passed to the engine, only stored in the output;
• had no cross-frame parallelism (serial loop);
• was not the MATLAB segment-based path the production pipeline uses.

Fix: rewrote track() to mirror StepDWorkflow::runNcorrAnalysis (the proven stepD path), but driven from in-memory cv::Mat frames (no temp-PNG round trip). It now builds one in-memory Image2D stack (ref + all frames), sets update_corrcoef to disable correlation-based reference updates, passes the user seed, and runs a single matlab_DIC_analysis_parallel / _sequential (or the exact_matlab_* variants) pass — honouring config.parallel_processing and config.ncorr_use_exact_matlab.

This brings in all three optimizations the production engine uses: parallel run, segment-based matlab-style, seeded ncorr. Semantically equivalent to the old per-frame loop (each frame still tracked against the fixed reference under NO_UPDATE), just far faster.

2. proxyncorr — .mat ROI/seed + all ncorr modes

• .mat ROI: --roi/roi= now accepts an image (PNG/BMP/…) or an xDIC .mat mask (refmask var), auto-detected by extension; default ROI now prefers <folder>/roi.mat then roi.png. The --in-memory path is .mat-aware too.
• .mat seed: --seeds/seeds_file= now accepts a JSON array or a REF_SEED_*.mat (seed_point [x,y]).
• All ncorr modes via -m/--mode: auto, sequential, parallel (existing) plus matlab-parallel, matlab-sequential, exact-matlab-parallel, exact-matlab-sequential (mode strings normalised; help/config docs updated).
• CMake: add matio/HDF5 include+link to the proxyncorr target.

Testing

• Both targets build cleanly with -DBUILD_SINGLEDIC=ON -DBUILD_PROXYNCORR=ON under -Wall -Wextra (no new warnings from changed code).
• ⚠️ Runtime not verified on this machine: the local OpenCV (4.13) is linked against protobuf 33 while Homebrew has protobuf 35.1, so opencv_dnn fails to load with an ABI mismatch — this breaks every binary equally (the pre-existing cppxdic --help fails identically). It's a local install skew, unrelated to these changes. Reviewer with a working OpenCV/protobuf should smoke-test:
  • singledic on a single-camera trial (compare wall-clock per frame to before),
  • proxyncorr -m matlab-parallel --roi REF_MASK_*.mat --seeds REF_SEED_*.mat.

🤖 Generated with Claude Code