scenescape-setup

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

SceneScape End-to-End Setup

SceneScape 端到端部署设置

Host needs Docker, docker-compose, and Python 3.10+ with

requests

主机需要安装Docker、docker-compose以及带有

requests

库的Python 3.10+。

Overview

概述

This skill deploys and resumes an Intel® SceneScape environment outside the repo, gathers the required deployment inputs from the user, and orchestrates the bootstrap, calibration, scene reconstruction, and verification workflow. It is intended for first-time installs, re-runs with existing

deploy-inputs.json

, and targeted phase resumes such as

bootstrap

calibrate

, or

scene

when the user only needs to repeat or continue a part of the deployment.

本技能用于在代码库外部部署和恢复Intel® SceneScape环境，收集用户提供的部署所需输入，并编排引导、校准、场景重建和验证工作流。适用于首次安装、使用现有

deploy-inputs.json

重新运行，以及当用户仅需重复或继续部署的某一部分时，针对性地恢复特定阶段（如

bootstrap

、

calibrate

或

scene

）。

Parameters / Arguments

参数/参数说明

Required runtime inputs for a fresh deployment:

deploy_dir

streams

(or video files),

camera_ids

scene_name

mapping

(scene map source:

reconstruction

default, blueprint,

.glb

.ply

mesh, or geospatial). Optional state fields:

--phase

--fresh

, and the resume flag implied by the Fast Path.

全新部署所需的运行时输入：

deploy_dir

、

streams

（或视频文件）、

camera_ids

、

scene_name

、

mapping

（场景地图来源：默认

reconstruction

，蓝图、

.glb

.ply

网格或地理空间数据）。可选状态参数：

--phase

、

--fresh

，以及快速路径隐含的恢复标志。

Returns / Output

返回结果/输出

Deployment artifacts in

deploy_dir

deploy-inputs.json

(source of truth),

.deploy-state.json

, orchestrator logs, calibration/reconstruction/verification outputs, and a final

DEPLOY COMPLETE

with a

scene_uid

and deployment metrics.

deploy_dir

中的部署产物：

deploy-inputs.json

（真实数据源）、

.deploy-state.json

、编排器日志、校准/重建/验证输出，以及包含

scene_uid

和部署指标的最终

DEPLOY COMPLETE

提示。

Error handling

错误处理

Fail safely instead of guessing: mismatched/duplicate streams vs

camera_ids

→ stop and ask for corrected inputs; unreadable prior inputs on a camera-change fresh redeploy → ask the user to confirm the retained set; missing local repo/docs → fall back to the canonical GitHub URL rather than fabricating; resume/continue signal → treat

deploy-inputs.json

as existing and skip Step 1 unless the user says the directory is wrong; a failed step → read only the matching troubleshooting reference, no broad log dumps.

安全失败而非猜测：流与

camera_ids

不匹配/重复→停止并要求用户提供修正后的输入；更换摄像头重新部署时无法读取先前输入→请用户确认保留的设置；本地代码库/文档缺失→使用官方GitHub URL作为替代而非自行编造；收到恢复/继续信号→将

deploy-inputs.json

视为已存在并跳过步骤1，除非用户指出目录错误；步骤失败→仅读取匹配的故障排除参考文档，不输出大量日志。

File resolution

文件解析

All scripts, references, and assets resolve relative to

$SKILL_DIR

, so the skill folder is self-contained and portable.

docs/user-guide/...

links point at the local checkout first; if unavailable (standalone skill copy), fall back to

https://github.com/open-edge-platform/scenescape/blob/main/<path>

instead of guessing. Never copy SceneScape repo docs into

references/

; reserve new references for knowledge that has no written form elsewhere.

所有脚本、参考文档和资源均相对于

$SKILL_DIR

解析，因此本技能文件夹是独立且可移植的。

docs/user-guide/...

链接首先指向本地检出的文件；若不可用（独立技能副本），则使用

https://github.com/open-edge-platform/scenescape/blob/main/<path>

作为替代而非猜测。请勿将SceneScape代码库中的文档复制到

references/

目录；仅将无书面记录的知识添加为新参考文档。

Always-on rules (no exceptions)

通用规则（无例外）

Before any deploy/resume/phase launch, read agent-guardrails.md.
Every orchestrator launch also starts
```
watch_orchestrator.sh
```
on the orchestrator PID in the background, notifying on
```
RESULT=
```
; never ask the user to poll status.
Never invent camera IDs/streams/scene names; never interpolate raw inputs into ad hoc shell one-liners; destructive actions (
```
--fresh
```
, deleting
```
deploy_dir
```
,
```
docker compose down -v
```
) always need explicit confirmation.
Load only the single phase/symptom reference that matches a reported failure.

在执行任何部署/恢复/阶段启动操作前，请阅读agent-guardrails.md。
每次启动编排器时，都会在后台针对编排器PID启动
```
watch_orchestrator.sh
```
，并在
```
RESULT=
```
时发出通知；绝不要求用户轮询状态。
切勿自行编造摄像头ID/流/场景名称；切勿将原始输入插入临时Shell单行命令；破坏性操作（
```
--fresh
```
、删除
```
deploy_dir
```
、
```
docker compose down -v
```
）始终需要用户明确确认。
仅加载与报告故障匹配的单个阶段/症状参考文档。

Step 0 — Bootstrap skill-dir

步骤0 — 引导技能目录

Resolve

SKILL_DIR

before any other step, using the first matching strategy:

A. Scripts already on disk (scenescape repo is checked out locally):

bash

export SKILL_DIR=<path-to-scenescape-checkout>/.github/skills/scenescape-setup

B. Extract from git (no full checkout needed — fast, leaves no branch state):

bash

SCENESCAPE_REPO=$(find ~ -maxdepth 5 -type d -name scenescape 2>/dev/null | head -1)
git -C "$SCENESCAPE_REPO" fetch origin main
mkdir -p /tmp/scenescape-skill
git -C "$SCENESCAPE_REPO" archive origin/main \
  -- .github/skills/scenescape-setup | tar -x -C /tmp/scenescape-skill
export SKILL_DIR=/tmp/scenescape-skill/.github/skills/scenescape-setup

Verify:

ls "$SKILL_DIR/scripts/deploy_scenescape.sh"

must succeed before continuing.

在执行任何其他步骤前，使用以下首个匹配策略解析

SKILL_DIR

：

A. 脚本已存在于本地磁盘（已本地检出scenescape代码库）：

bash

export SKILL_DIR=<scenescape检出路径>/.github/skills/scenescape-setup

B. 从Git提取（无需完整检出——速度快，不会留下分支状态）：

bash

SCENESCAPE_REPO=$(find ~ -maxdepth 5 -type d -name scenescape 2>/dev/null | head -1)
git -C "$SCENESCAPE_REPO" fetch origin main
mkdir -p /tmp/scenescape-skill
git -C "$SCENESCAPE_REPO" archive origin/main \
  -- .github/skills/scenescape-setup | tar -x -C /tmp/scenescape-skill
export SKILL_DIR=/tmp/scenescape-skill/.github/skills/scenescape-setup

验证：在继续操作前，

ls "$SKILL_DIR/scripts/deploy_scenescape.sh"

命令必须执行成功。

Routing

场景路由

Situation	Reference to read
New deployment (gather inputs, mapping choice, video files)	step-1-gather-inputs.md
Resume / repeat / Fast Path ("continue", "resume", unchanged inputs)	fast-path.md
Launch (full deploy, resume, or `--phase` orchestrator + watcher + README + handoff)	deploy-and-complete.md
Single phase: bootstrap (6–8), calibrate (9–10), scene (11–13)	phase-bootstrap.md / phase-calibrate.md / phase-scene.md
Tracking flickers, vanishes, or IDs change (same camera)	tuning-tracker.md
Cross-camera Re-ID misses / wrong person	tuning-reid.md
Keep a vision attribute from resetting	attribute-persistence.md
External non-vision sensor reading/event	singleton-sensors.md
Expected size/shape for a class (Object Library)	object-library.md
After successful deploy — what to build with scene output (required handoff)	using-scene-output.md
Generated-file layout / web-UI handoff / bootstrap-runtime-reconstruction diagnosis	operational-reference.md (only for those needs — not during routine deploy)

场景	需阅读的参考文档
全新部署（收集输入、映射选项、视频文件）	step-1-gather-inputs.md
恢复/重复/快速路径（“继续”、“恢复”、输入未更改）	fast-path.md
启动（完整部署、恢复或 `--phase` 编排器 + 监视器 + README + 交接）	deploy-and-complete.md
单个阶段：引导（6–8）、校准（9–10）、场景（11–13）	phase-bootstrap.md / phase-calibrate.md / phase-scene.md
跟踪闪烁、消失或ID变化（同一摄像头）	tuning-tracker.md
跨摄像头Re-ID识别错误/匹配错误人员	tuning-reid.md
保持视觉属性不被重置	attribute-persistence.md
外部非视觉传感器读取/事件	singleton-sensors.md
类别的预期尺寸/形状（对象库）	object-library.md
部署成功后——如何使用场景输出（必需交接）	using-scene-output.md
生成文件布局 / Web UI交接 / 引导-运行时-重建诊断	operational-reference.md（仅用于这些需求——常规部署期间无需查看）

Tuning tracker/Re-ID behavior (reactive only)

跟踪器/Re-ID行为调优（仅响应式操作）

Do not ask tuning questions upfront during Step 1 — always deploy with the shipped

tracker-config.json

reid-config.json

defaults first. Open the matching questionnaire only after the user reports tracking/Re-ID dissatisfaction. In that first response:

State which reference you opened (
```
tuning-tracker.md
```
or
```
tuning-reid.md
```
— exactly one).
Present that reference's numbered questionnaire in your reply.
In the same turn, apply symptom-derived starter values from that reference's recommendation logic to the deployed copy at
```
<deploy_dir>/controller/tracker-config.json
```
or
```
<deploy_dir>/controller/reid-config.json
```
(never the skill's
```
assets/
```
originals). Show the exact JSON field changes and the exact restart command
```
docker compose up -d --force-recreate scene
```
.
Note that questionnaire answers can further refine the starter values.

Do not skip the questionnaire, and do not skip showing the deployed-path edits + scene-only restart. Load exactly one matching reference (tracker timing vs cross-camera Re-ID).

在步骤1中切勿提前询问调优问题——始终先使用默认的

tracker-config.json

reid-config.json

配置进行部署。仅在用户报告跟踪/Re-ID不满意之后，才打开匹配的调查问卷。首次响应时需执行以下操作：

说明你打开了哪个参考文档（
```
tuning-tracker.md
```
或
```
tuning-reid.md
```
——仅选其一）。
在回复中展示该参考文档中的编号调查问卷。
在同一轮对话中，根据参考文档推荐逻辑中的症状推导初始值，修改部署副本中的
```
<deploy_dir>/controller/tracker-config.json
```
或
```
<deploy_dir>/controller/reid-config.json
```
（切勿修改技能
```
assets/
```
目录中的原始文件）。展示具体的JSON字段变更以及精确的重启命令
```
docker compose up -d --force-recreate scene
```
。
说明调查问卷的答案可进一步优化初始值。

请勿跳过调查问卷，也请勿跳过展示部署路径的修改内容 + 仅重启场景服务的步骤。仅加载一个匹配的参考文档（跟踪时序问题对应跨摄像头Re-ID问题）。

Quality & Evaluation

质量与评估

Automated eval cases live in evals/evals.json, one entry per

example-prompts/

file (

prompt_file

links the two together). See benchmark/benchmark.md for the current benchmark.

自动化评估案例位于evals/evals.json中，每个

example-prompts/

文件对应一个条目（

prompt_file

字段将两者关联）。当前基准测试详情请查看benchmark/benchmark.md。