跳到正文
cccskills
免费在 GitHub 公开

depth-inspect

Inspect native RGB, depth, confidence, calibration, and poses to make finite, source traceable reference points and surface measurements.

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(1)

  • SKILL.md4.3 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

Use when depth availability, orientation, or surface support is in question. Record RGB/depth/confidence/calibration orientation and selectable raw/clean depth and raw/corrected pose variants. Distinguish depth availability from local repeatability and absolute accuracy. Flag glass background returns, occluded furniture, opaque surfaces without support, and nonfinite samples.

Segment physical patches before fitting: ramp, floor, curb, soffit, jamb return, and other planes are separate surfaces. Retain rejected samples, all point residuals, and inlier statistics. Read formats.md and evidence-workflow.md.

Use scripts/surfaces.py with the exact filename stem, image orientation, physically reviewed polygon, surface identity and selected variants. Declare the capture's tracking segment when its camera metadata supplies one, especially before applying a registered transform; the declaration must match the native raw camera rather than a filename or corrected-pose inference. Preserve missing segment metadata as unknown instead of substituting segment 0. Inspect both the photo polygon and native depth map in its output. Reject proposals that include a door, object or another plane even if their labels sound correct. Do not convert a contact-sheet index into a timestamp by estimation; look it up in the index.

Each measured patch reports native depth zero/nonzero, in-range/out-of-range and confidence-value counts over the selected mask. These counts describe encoded availability and filtering only; they do not establish physical support or sensor accuracy. With --pixel-inspections, the source crop also marks the eligible native depth sample centers after the existing depth-to-RGB mapping and records that mapping for pending visual review.

Before running a measurement, bind the exact reviewed selection record and serialized spec hashes. Programmatically compare the frame/source identity, orientation and polygon between the reviewed selection and the spec, then verify those same fields in the output. A revised preview does not authorize reusing an earlier mask whose serialized selection was not updated.

To review mapped sample locations before fitting, set "fit": false on every patch and "comparisons": [] in the spec. Inspect the retained eligible native pixels and points against the source; this mode does not validate the selection's physical meaning. Its all-false inlier_mask means not fitted, not rejected samples. Enable fitting only in a fresh run after physical sample review; see formats.md.

For small patches or ambiguous edges, add --pixel-inspections. Inspect the complete unmarked/marked polygon crop and every vertex's native-pixel context before interpreting support. Nearest-neighbor enlargement adds no detail, and a clean vertex does not prove the polygon interior avoids another surface. These optional outputs leave sampling and fitting unchanged. A capable agent can perform this visual review by opening the actual images and recording its findings; generated labels, counts or successful commands are not a review. Keep uninspected or ambiguous selections pending. Seek human input for an explicit owner checkpoint or a necessary physical fact the available evidence cannot resolve, rather than requiring a human for every patch. Reject a mixed-surface selection and review its correction before fitting.

Compare the same physical patch only within a shared coordinate frame and overlapping footprint. Reserve sufficiently separated views and independent features to expose pose/depth correlation. Report support counts, all-sample and retained residuals, overlap size and cross-view disagreement separately. Measure a slope freely before deciding whether it should be horizontal. Do not call a mixed ramp/floor fit a scan error.

If a boundary lacks native depth, report it as an image observation or an intersection inferred from identified planes. Do not give that boundary the same confidence as a directly supported patch. An optional physical length can check absolute scale; it is not a prerequisite for processing calibrated native depth.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

Build and review evidence tagged Blender architectural candidates, preserve unrelated scene content, and produce requested photographic materials, stills and portable model deliveries.

中文简介正在准备中。当前显示原文说明。

obra/scan-to-model72026年9月27日 更新

Preserve and inventory scan archives while retaining capture variants, hashes, calibration, and frame provenance.

中文简介正在准备中。当前显示原文说明。

obra/scan-to-model72026年9月27日 更新

Fit and independently validate rigid scan alignment from named, source traceable correspondences.

中文简介正在准备中。当前显示原文说明。

obra/scan-to-model72026年9月27日 更新

Coordinate evidence based reconstruction from scans, photos, plans, and measurements into a reviewable Blender architectural model, with photographic appearance, renders and portable browser delivery when requested.

中文简介正在准备中。当前显示原文说明。

obra/scan-to-model72026年9月27日 更新

Use when cataloging or annotating visible construction in photographs or scan RGB, especially before registration or while the base model may be replaced.

中文简介正在准备中。当前显示原文说明。

obra/scan-to-model72026年9月27日 更新

Collect and verify Google Street View exterior references for a building reconstruction, preserving panorama identity, imagery dates, attribution and viewpoint separately from calibrated scan evidence.

中文简介正在准备中。当前显示原文说明。

obra/scan-to-model72026年9月27日 更新

查看 obra 的全部技能

报告此技能的问题