imolodetskikh/sr-artifact-prominence

SR Artifact Prominence

Annotated super-resolution artifact regions across four image subsets, with crowdsourced per-region prominence scores, artifact type labels, and natural-language descriptions.

Prominence is the fraction of valid crowd workers who answered that the highlighted region contains a noticeable super-resolution artifact.

Subsets

Subset	Source dataset	Source images	Masks	Notes
open_images	Open Images	547	1,523	GT + LR-bicubic + multiple SR variants
open_images_uncurated	Open Images	316	826	Subset of open_images used for the uncurated comparison and scoring; uses the same images/open_images/ tree
urban100	Urban100	68	873	GT + LR-bicubic + multiple SR variants
urban100_hr	Urban100	94	946	Original Urban100 images used directly as SR inputs; no higher-resolution GT
desra	DeSRA dataset	528	593	Annotations only

Layout

annotations/
  {open_images,urban100,urban100_hr,desra}.csv
images/
  open_images/<folder>/<files>
  urban100/<folder>/<files>
  urban100_hr/<folder>/<files>

For each annotated folder, the image directory contains:

• <folder>.png — source/reference image (HR ground truth for open_images and urban100; the original Urban100 image used as the SR input for urban100_hr)
• <folder>@LR@bicubic_x4.png — bicubic LR input (open_images and urban100 only)
• <folder>@SR@RLFN_x4.png — RLFN SR output, included as a reference SR where available
• All sr_fn and mask_fn files referenced by rows of the corresponding CSV

mask_fn files (@MS@dist.png) are binary candidate artifact masks shown to workers. The scalar prominence score for each mask is stored in the CSV as ratio_bi.

CSV columns

• folder, sr_fn, mask_fn — image identifiers
• ratio_bi — crowdsourced prominence: fraction of valid workers who confirmed an artifact in the highlighted region
• dilated — whether mask preprocessing was applied before crowdsourcing (morphological opening, dilation, and closing for easier visual assessment); absent for desra
• type — artifact-type label (e.g., hallucinated texture, plastic texture)
• description — natural-language description of the artifact
• semantic_categories, semantic_description — scene-content tags
• status, semantic_status — annotation completion flags

DeSRA subset

We do not redistribute DeSRA images. The desra.csv rows reference the original DeSRA filenames (<folder> and <sr_fn> columns), which derive from ImageNet (subject to the ImageNet license). Obtain the images from the DeSRA project release and join on folder + sr_fn.

DeSRA: Xie et al., "DeSRA: Detect and Delete the Artifacts of GAN-based Real-World Super-Resolution Models," ICCV 2023.

Heatmaps and reproducibility

heatmaps/<subset>/<gt_dir>/<method>/<sr_stem>.npy.gz contains per-pixel artifact-prominence heatmaps. Each heatmap is a gzipped 2-D NumPy array aligned to the SR image named by sr_fn; the file stem matches that image stem (for example, img@SR@RLFN_x4.png corresponds to img@SR@RLFN_x4.npy.gz).

The <gt_dir> level records which reference image the method was scored against:

Subset	<gt_dir>	Reference
open_images	gt_RLFN_x4	RLFN pseudo-GT
urban100	gt_GT	High-resolution GT

Methods shipped here (limited due to dataset size concerns):

• bd_jup — block feature used by the paper's reference baseline
• dists — the DISTS full-reference metric
• desra — the DeSRA artifact detector
• nn-20250421-gtgt-e30 — the paper's shallow neural baseline that fuses bd_jup, ssm_jup, and DISTS-derived features

Released heatmaps cover the open_images and urban100 subsets. We do not redistribute heatmaps for urban100_hr because of size, or for desra because the underlying DeSRA/ImageNet images are not redistributed here. For no-HR evaluation, the paper uses the RLFN pseudo-GT protocol described in the paper.

contrast50.py reproduces the paper's threshold-free prominence score (contrast_p50:Spearman):

python contrast50.py \
    --dataset-root . \
    --subset open_images \
    --heatmap-dir heatmaps/open_images/gt_RLFN_x4/nn-20250421-gtgt-e30

For each annotated mask, it computes median(heatmap[mask]) - median(heatmap[~mask]) and reports the Spearman correlation with ratio_bi over the selected subset. Rows whose masks were marked dilated=True are eroded before scoring, matching the paper evaluation.

For open_images, the paper reports contrast scores on the uncurated subset. Reproduce with --annotations-csv:

python contrast50.py \
    --dataset-root . \
    --subset open_images \
    --annotations-csv annotations/open_images_uncurated.csv \
    --heatmap-dir heatmaps/open_images/gt_RLFN_x4/nn-20250421-gtgt-e30

Licensing

• All annotations (every CSV in annotations/): CC-BY 4.0.
• open_images images: CC-BY 2.0, inherited from Open Images.
• urban100, urban100_hr images: CC, inherited from Urban100.
• desra images: not redistributed here; the DeSRA images derive from the ImageNet-1K validation set and are governed by the ImageNet terms of access (non-commercial research and educational use).