Benchmarking inpainting methods for removing bird/insect protection netting from photos. The test photos are all shot through a fine mesh net that spans the entire frame (rooftop/balcony shots looking out through the net), not a localized fence or sports net — that shapes the whole approach below. See note.md for the original, broader implementation plan.
Stage 1 — mask generation (generate_masks.py):
classical CLAHE + black-hat morphological thresholding, run per image. SAM 2
(point/box-prompt object segmentation) doesn't fit this dataset — there's no
discrete object to prompt, just a uniform grid texture over the whole photo —
so a background-adaptive classical detector is used instead. Masks are saved
to masks/.
Stage 2 — inpainting (methods/iopaint_methods.py):
runs the frozen mask through three methods via IOPaint's ModelManager:
- Telea (OpenCV, classical, no GPU)
- LaMa (FFC architecture, GPU)
- MI-GAN (lightweight GAN, GPU, run at its native 512×512)
Each method's output gets a light median-filter + unsharp-mask cleanup before compositing back with the original (unmasked pixels always come from the untouched original image, never from resized/inpainted output).
SDXL-inpaint was also tried and dropped: ~24s/image versus sub-second for the other three, for no clear quality win on this kind of thin repeating structure.
Orchestration (run_benchmark.py): loops over every
data/*.JPG, runs all three methods, and saves a per-image 2×3 comparison
figure (results/<name>_comparison.png) plus one overview grid
(results/overview_grid.png).
make setup # venv + dependencies
.venv/bin/python3 generate_masks.py
.venv/bin/python3 run_benchmark.py
Rows are Original / Telea / LaMa / MI-GAN; columns are the individual test photos.
Telea is the best performer — by visual inspection (repo owner's judgment), the plain OpenCV classical method produces the cleanest result of the three on this dataset. LaMa and MI-GAN don't clearly outperform it here, despite being the more sophisticated, GPU-based methods. Telea also happens to be the cheapest option (CPU-only, no model weights).
This somewhat inverts the usual expectation (learned methods beating classical baselines) and is likely specific to this dataset's structure: a thin, high-frequency, repeating pattern over mostly smooth/low-detail backgrounds (sky, clouds, flat concrete) is close to the textbook case Telea's fast marching method was designed for, whereas LaMa and MI-GAN are tuned more for larger, irregular holes with structured content around them.
- The classical mask (bottom-left panel in each comparison figure) is clean over sky/cloud but noisier over foliage and can miss the net entirely in the darkest regions — room to improve before drawing firm conclusions.
- Inpainting runs at a capped working resolution (1536px long side, or 512x512 for MI-GAN) rather than native resolution, for speed.
- No quantitative metrics yet (BRISQUE/NIQE/PSNR/SSIM/LPIPS) — this has been a qualitative comparison so far.
- FLUX.1-Fill could be tried as a stronger diffusion contender than SDXL-inpaint (see note.md) — it was left out of this pass to keep things fast and disk-light, not ruled out on quality grounds.