[Bug] Issue with Python port of Liu et al.

[EnoxmolingSoftware]

Method

Chou & Li

Language port

MATLAB

Describe the bug

I cloned the repo and ran the Liu et al. Python port on a grayscale image by editing the hardcoded path in Liu et al/Python/main.py:

cd "Liu et al/Python"
# edit main.py: distImg = cv2.imread('<my image>.bmp', cv2.IMREAD_GRAYSCALE)
python main.py

The output JND map looks qualitatively reasonable (low at edges, high in textures), but it doesn't agree with the MATLAB port on the same image — textured regions get a visibly higher budget in Python than in MATLAB. Looking at main.py I see jnd_id uses cv2.GaussianBlur(I, (7, 7), sigma) to do the structural / textural split, while my MATLAB run of JND_ID(im, 1.0) clearly calls something else.

Is the Python jnd_id(I, lambda_=0.8) meant to be equivalent to the MATLAB JND_ID(I, lambda), or a deliberate simplification? Tried sweeping lambda_ from 0.1 to 2.0 — the gap stays.

Minimal reproduction

.

Environment

Additional context

No response

[Terriao]

Hi EnoxmolingSoftware, not a bug — this is a deliberate simplification documented in the README, but it's worth restating here for visibility.

The MATLAB JND_ID(I, lambda) performs the structural / textural decomposition with Wotao Yin's ParaMaxFlow TV-L¹ solver, which ships as a 64-bit Windows MEX. To keep the Python port runnable on Linux / macOS with no MATLAB or MEX dependency, jnd_id(I, lambda_) substitutes a Gaussian-blur surrogate (the cv2.GaussianBlur(I, (7, 7), sigma) you spotted) for that decomposition. The README Runtime analysis section notes this and the consequence — "Where we deliberately simplified a costly step (TV-L¹ in Liu) we say so explicitly."

The qualitative ranking (textures > edges, larger budget where the eye masks more) is preserved by design, but the absolute amplitudes in textured regions legitimately differ — which is exactly what you're seeing. lambda_ controls the blur sigma, not the TV regulariser, so sweeping it can't close the gap.

Two practical options:

1. Use the MATLAB port on Windows if you need bit-faithful Liu et al. output for a downstream codec.
2. Stay in Python and post-calibrate. For IQA / watermarking, where only relative budgets matter, a per-image affine fit against a few MATLAB-rendered maps is usually enough. For codec-driving, where absolute budgets matter, option 1 is safer.

A pure-Python TV-L¹ that closes the gap without the MEX would be the right long-term fix and a great contribution.

feel free to reopen if the qualitative behaviour still looks off after re-reading the README's Runtime analysis note.