ADE20K Texture Miner: Research Idea, Design, and Ranking Evidence
The core research idea behind this miner was to treat texture-rich scene discovery as a ranking problem rather than a binary classifier from the start. The practical question was whether a large natural-image corpus like ADE20K can be ordered by a texture-awareness signal that is meaningful for downstream texture research, without hand-labeling a new dataset. Instead of asking a model for object semantics first, the miner asks for region structure, boundary behavior, and texture occupancy signals, then fuses them into a final ranking score intended to push highly textured images upward and semantically dominant but weak-texture scenes downward.
The research question was explicit: can an automatically computed score preserve a useful monotonic order of texture richness at scale? The design therefore separated evidence collection and decision logic. First, the pipeline extracted region proposals and measured structural statistics that are expected in textured scenes: broader texture occupancy, stronger internal boundary structure, and multiple coherent texture regions. Second, it integrated semantic safeguards to avoid promoting object-centric images that can look visually busy but are not texture-rich in the sense needed for texture segmentation work. Third, it fused the evidence into one final score on a 0-100 scale and kept all component metrics for auditing so ranking behavior could be inspected instead of treated as a black box.
In practical terms, the solution was implemented as a staged miner with explicit intermediate signals and a final sortable manifest. The output is not only a top list; it is a full-ranked set with per-image component scores, status labels, masks, overlays, and original thumbnails. That design choice matters because it allows you to validate the ranking logic by checking whether high-score images consistently exhibit stronger texture-centric geometry than low-score images. The sections below present that validation directly from the 1,500 ADE20K entries currently hosted in this standalone site.
Ranking Summary
These statistics are computed from the live ADE20K review manifest. The top-vs-bottom comparison uses the highest and lowest 10% by final rank score.
Evidence That Ranking Tracks Texture Richness
Decile Trend Across Rank
Decile 1 is highest-ranked. Lines are normalized per metric to expose trend shape, not absolute units.
Scatter: Final Score vs Texture Fraction
Each point is one image. Regression line should rise if ranking aligns with texture occupancy.
Scatter: Final Score vs Object Fraction
A texture-aware rank should suppress heavy object-dominant scenes, producing a negative slope.
Top 10% vs Bottom 10%
Mean metric comparison between top-ranked and bottom-ranked subsets.
Rank Explorer
Move through the global ranking and inspect the image plus its score decomposition.
Visual Rank Extremes
Top examples should visually present stronger texture transitions and texture occupancy, while bottom examples should show weaker texture evidence or stronger object dominance.
Top 12
Bottom 12
Top-50 Ranked Table
| Rank | Image ID | Final Score | Texture Fraction | Object Fraction | Boundary Score | Status |
|---|