Clarification / release of frozen single-block ViT corner detector used in pretraining

[georgeyiasemis]

Summary

Sec. 3.4 of the paper describes online boundary target generation that depends on a frozen single-block Vision Transformer to seed sparse corner points. The public repo appears to ship backbone inference + PCA demo only, with no corner detector code or weights.

Questions

1. Will the corner-point ViT (architecture + checkpoint) be released?
2. What was it trained on (dataset / labels / objective)? The paper does not specify this.
3. If it will not be released, what is the recommended substitute for reproduction (e.g. field-derived corners $$j_\mathrm{loc}=1-d $$ from Fig. 11, classical detectors, etc.)?

Context from the paper

• Corners anchor vote-aggregation decoding of the boundary field (Finding 1).
• Described as “orders of magnitude smaller than the backbone” and never observing boundary fields.
• Stated as the only fixed component outside the self-supervised backbone loop.