gaojie
dbcbdbb59e
- 处理: 74 个 .md 文件 - 跳过: 0 个（无已存在的 front matter） - 异常: 3 个（H1 缺失，用文件名兜底） - plans/PRISM/.research/readmes/3d-llm.md - plans/PRISM/.research/readmes/openmask3d.md - plans/PRISM/.research/readmes/openscene.md
700 lines
22 KiB
Markdown
700 lines
22 KiB
Markdown
---
|
||
title: "酒店场景室内建模与物理验证项目详细实施计划（续）"
|
||
date: 2026-05-20
|
||
draft: false
|
||
tags: ["规划", "酒店场景", "物理"]
|
||
categories: ["worldmodel"]
|
||
---
|
||
|
||
# 酒店场景室内建模与物理验证项目详细实施计划（续）
|
||
|
||
## 七、数据管理与存储架构（续）
|
||
|
||
### 7.2 数据库设计
|
||
|
||
```sql
|
||
-- PostgreSQL + PostGIS空间数据库
|
||
|
||
-- 场景表
|
||
CREATE TABLE scenes (
|
||
scene_id SERIAL PRIMARY KEY,
|
||
scene_type VARCHAR(50), -- 'lobby', 'corridor', 'room', 'bathroom'
|
||
hotel_name VARCHAR(100),
|
||
floor_number INTEGER,
|
||
capture_date TIMESTAMP,
|
||
sensor_config JSONB,
|
||
bbox_min GEOMETRY(POINTZ, 4326),
|
||
bbox_max GEOMETRY(POINTZ, 4326),
|
||
metadata JSONB
|
||
);
|
||
|
||
-- 物体实例表
|
||
CREATE TABLE object_instances (
|
||
instance_id SERIAL PRIMARY KEY,
|
||
scene_id INTEGER REFERENCES scenes(scene_id),
|
||
label VARCHAR(100),
|
||
category VARCHAR(50),
|
||
centroid GEOMETRY(POINTZ, 4326),
|
||
bbox_min GEOMETRY(POINTZ, 4326),
|
||
bbox_max GEOMETRY(POINTZ, 4326),
|
||
volume FLOAT,
|
||
mesh_path TEXT,
|
||
semantic_features VECTOR(512), -- CLIP特征，使用pgvector扩展
|
||
attributes JSONB, -- {material, color, state, affordance}
|
||
confidence FLOAT
|
||
);
|
||
|
||
-- 空间关系表
|
||
CREATE TABLE spatial_relations (
|
||
relation_id SERIAL PRIMARY KEY,
|
||
source_instance_id INTEGER REFERENCES object_instances(instance_id),
|
||
target_instance_id INTEGER REFERENCES object_instances(instance_id),
|
||
relation_type VARCHAR(50), -- 'supported_by', 'inside', 'next_to', 'above'
|
||
confidence FLOAT,
|
||
metadata JSONB
|
||
);
|
||
|
||
-- 重建模型表
|
||
CREATE TABLE reconstruction_models (
|
||
model_id SERIAL PRIMARY KEY,
|
||
scene_id INTEGER REFERENCES scenes(scene_id),
|
||
model_type VARCHAR(50), -- '3dgs', 'nerf', 'mesh'
|
||
file_path TEXT,
|
||
quality_metrics JSONB, -- {psnr, ssim, lpips}
|
||
training_config JSONB,
|
||
created_at TIMESTAMP
|
||
);
|
||
|
||
-- 物理属性表
|
||
CREATE TABLE physics_properties (
|
||
property_id SERIAL PRIMARY KEY,
|
||
instance_id INTEGER REFERENCES object_instances(instance_id),
|
||
mass FLOAT,
|
||
friction_static FLOAT,
|
||
friction_dynamic FLOAT,
|
||
restitution FLOAT,
|
||
is_articulated BOOLEAN,
|
||
joint_type VARCHAR(50), -- 'revolute', 'prismatic', 'fixed'
|
||
joint_params JSONB
|
||
);
|
||
|
||
-- 创建空间索引
|
||
CREATE INDEX idx_scenes_bbox ON scenes USING GIST(bbox_min);
|
||
CREATE INDEX idx_instances_centroid ON object_instances USING GIST(centroid);
|
||
CREATE INDEX idx_instances_semantic ON object_instances USING ivfflat(semantic_features vector_cosine_ops);
|
||
```
|
||
|
||
### 7.3 数据存储方案
|
||
|
||
```yaml
|
||
存储层级:
|
||
热数据（频繁访问）:
|
||
- 存储: NVMe SSD RAID 10
|
||
- 容量: 4TB
|
||
- 内容:
|
||
- 当前处理中的原始数据
|
||
- 训练好的3DGS/NeRF模型
|
||
- 数据库文件
|
||
- 备份: 每日增量备份
|
||
|
||
温数据（偶尔访问）:
|
||
- 存储: SATA HDD RAID 6
|
||
- 容量: 20TB
|
||
- 内容:
|
||
- 历史原始数据
|
||
- 中间处理结果
|
||
- 渲染缓存
|
||
- 备份: 每周全量备份
|
||
|
||
冷数据（归档）:
|
||
- 存储: 对象存储（MinIO/S3）
|
||
- 容量: 无限扩展
|
||
- 内容:
|
||
- 完成项目的完整数据集
|
||
- 多版本模型
|
||
- 实验日志
|
||
- 备份: 异地容灾
|
||
|
||
对象存储结构:
|
||
bucket: hotel-reconstruction
|
||
├── raw-data/
|
||
│ └── {hotel_name}/{scene_type}/{date}/
|
||
├── processed/
|
||
│ └── {hotel_name}/{scene_type}/{version}/
|
||
├── models/
|
||
│ └── {model_type}/{scene_id}/{checkpoint}/
|
||
└── exports/
|
||
└── {format}/{scene_id}/
|
||
```
|
||
|
||
### 7.4 数据版本控制
|
||
|
||
```yaml
|
||
使用DVC（Data Version Control）:
|
||
|
||
初始化:
|
||
$ dvc init
|
||
$ dvc remote add -d storage s3://hotel-reconstruction
|
||
|
||
跟踪大文件:
|
||
$ dvc add data/rooms/room_301/raw/
|
||
$ git add data/rooms/room_301/raw/.dvc
|
||
$ git commit -m "Add room 301 raw data"
|
||
|
||
版本切换:
|
||
$ git checkout v1.0
|
||
$ dvc checkout
|
||
|
||
数据管道:
|
||
# dvc.yaml
|
||
stages:
|
||
slam:
|
||
cmd: python scripts/run_slam.py
|
||
deps:
|
||
- data/raw/
|
||
- scripts/run_slam.py
|
||
outs:
|
||
- data/processed/slam/
|
||
|
||
3dgs_training:
|
||
cmd: python scripts/train_3dgs.py
|
||
deps:
|
||
- data/processed/slam/
|
||
- scripts/train_3dgs.py
|
||
outs:
|
||
- models/3dgs/
|
||
metrics:
|
||
- metrics/3dgs_quality.json
|
||
```
|
||
|
||
---
|
||
|
||
## 八、评测基准与验证方案
|
||
|
||
### 8.1 几何重建质量评测
|
||
|
||
#### 8.1.1 点云精度评测
|
||
|
||
```python
|
||
class GeometryEvaluator:
|
||
def __init__(self, ground_truth_mesh, reconstructed_pointcloud):
|
||
self.gt_mesh = ground_truth_mesh
|
||
self.recon_pc = reconstructed_pointcloud
|
||
|
||
def compute_chamfer_distance(self):
|
||
"""Chamfer距离（双向最近点距离）"""
|
||
# GT mesh采样点云
|
||
gt_pc = self.gt_mesh.sample_points_uniformly(100000)
|
||
|
||
# 重建点云 → GT点云
|
||
dist_recon_to_gt = self.nearest_neighbor_distance(
|
||
self.recon_pc, gt_pc
|
||
)
|
||
|
||
# GT点云 → 重建点云
|
||
dist_gt_to_recon = self.nearest_neighbor_distance(
|
||
gt_pc, self.recon_pc
|
||
)
|
||
|
||
chamfer = (dist_recon_to_gt.mean() + dist_gt_to_recon.mean()) / 2
|
||
|
||
return {
|
||
'chamfer_distance': chamfer,
|
||
'recon_to_gt_mean': dist_recon_to_gt.mean(),
|
||
'gt_to_recon_mean': dist_gt_to_recon.mean(),
|
||
'recon_to_gt_std': dist_recon_to_gt.std()
|
||
}
|
||
|
||
def compute_accuracy_completeness(self, threshold=0.05):
|
||
"""准确率与完整性（阈值：5cm）"""
|
||
gt_pc = self.gt_mesh.sample_points_uniformly(100000)
|
||
|
||
dist_recon_to_gt = self.nearest_neighbor_distance(
|
||
self.recon_pc, gt_pc
|
||
)
|
||
dist_gt_to_recon = self.nearest_neighbor_distance(
|
||
gt_pc, self.recon_pc
|
||
)
|
||
|
||
# 准确率：重建点中有多少在GT附近
|
||
accuracy = (dist_recon_to_gt < threshold).mean()
|
||
|
||
# 完整性：GT点中有多少被重建覆盖
|
||
completeness = (dist_gt_to_recon < threshold).mean()
|
||
|
||
# F-score
|
||
f_score = 2 * accuracy * completeness / (accuracy + completeness + 1e-8)
|
||
|
||
return {
|
||
'accuracy': accuracy,
|
||
'completeness': completeness,
|
||
'f_score': f_score
|
||
}
|
||
|
||
def compute_normal_consistency(self):
|
||
"""法线一致性"""
|
||
# 估计重建点云的法线
|
||
recon_normals = self.estimate_normals(self.recon_pc)
|
||
|
||
# 对于每个重建点，找到GT mesh上最近的面
|
||
closest_faces = self.find_closest_faces(self.recon_pc, self.gt_mesh)
|
||
gt_normals = self.gt_mesh.face_normals[closest_faces]
|
||
|
||
# 计算法线夹角
|
||
dot_products = np.abs((recon_normals * gt_normals).sum(axis=1))
|
||
normal_consistency = dot_products.mean()
|
||
|
||
return normal_consistency
|
||
```
|
||
|
||
#### 8.1.2 渲染质量评测
|
||
|
||
```python
|
||
class RenderingEvaluator:
|
||
def __init__(self, model, test_cameras, ground_truth_images):
|
||
self.model = model
|
||
self.test_cameras = test_cameras
|
||
self.gt_images = ground_truth_images
|
||
|
||
def evaluate_all_metrics(self):
|
||
results = {
|
||
'psnr': [],
|
||
'ssim': [],
|
||
'lpips': [],
|
||
'rendering_time': []
|
||
}
|
||
|
||
for cam, gt_img in zip(self.test_cameras, self.gt_images):
|
||
# 渲染
|
||
start_time = time.time()
|
||
rendered_img = self.model.render(cam)
|
||
render_time = time.time() - start_time
|
||
|
||
# 计算指标
|
||
psnr = self.compute_psnr(rendered_img, gt_img)
|
||
ssim = self.compute_ssim(rendered_img, gt_img)
|
||
lpips = self.compute_lpips(rendered_img, gt_img)
|
||
|
||
results['psnr'].append(psnr)
|
||
results['ssim'].append(ssim)
|
||
results['lpips'].append(lpips)
|
||
results['rendering_time'].append(render_time)
|
||
|
||
# 统计
|
||
summary = {
|
||
'psnr_mean': np.mean(results['psnr']),
|
||
'psnr_std': np.std(results['psnr']),
|
||
'ssim_mean': np.mean(results['ssim']),
|
||
'lpips_mean': np.mean(results['lpips']),
|
||
'fps': 1.0 / np.mean(results['rendering_time'])
|
||
}
|
||
|
||
return summary
|
||
|
||
def compute_psnr(self, img1, img2):
|
||
"""峰值信噪比"""
|
||
mse = np.mean((img1 - img2) ** 2)
|
||
if mse == 0:
|
||
return float('inf')
|
||
return 20 * np.log10(1.0 / np.sqrt(mse))
|
||
|
||
def compute_ssim(self, img1, img2):
|
||
"""结构相似性"""
|
||
from skimage.metrics import structural_similarity
|
||
return structural_similarity(
|
||
img1, img2,
|
||
multichannel=True,
|
||
data_range=1.0
|
||
)
|
||
|
||
def compute_lpips(self, img1, img2):
|
||
"""感知相似性（使用预训练网络）"""
|
||
import lpips
|
||
loss_fn = lpips.LPIPS(net='alex')
|
||
|
||
# 转换为tensor
|
||
img1_t = torch.from_numpy(img1).permute(2, 0, 1).unsqueeze(0)
|
||
img2_t = torch.from_numpy(img2).permute(2, 0, 1).unsqueeze(0)
|
||
|
||
return loss_fn(img1_t, img2_t).item()
|
||
```
|
||
|
||
### 8.2 语义理解评测
|
||
|
||
```python
|
||
class SemanticEvaluator:
|
||
def __init__(self, predicted_instances, ground_truth_instances):
|
||
self.pred = predicted_instances
|
||
self.gt = ground_truth_instances
|
||
|
||
def compute_3d_iou(self):
|
||
"""3D IoU（Intersection over Union）"""
|
||
ious = []
|
||
|
||
for pred_inst in self.pred:
|
||
best_iou = 0
|
||
|
||
for gt_inst in self.gt:
|
||
if pred_inst['label'] != gt_inst['label']:
|
||
continue
|
||
|
||
# 计算3D包围盒IoU
|
||
intersection = self.bbox_intersection(
|
||
pred_inst['bbox'], gt_inst['bbox']
|
||
)
|
||
union = self.bbox_union(
|
||
pred_inst['bbox'], gt_inst['bbox']
|
||
)
|
||
|
||
iou = intersection / (union + 1e-8)
|
||
best_iou = max(best_iou, iou)
|
||
|
||
ious.append(best_iou)
|
||
|
||
return np.mean(ious)
|
||
|
||
def compute_map_3d(self, iou_threshold=0.5):
|
||
"""3D目标检测的mAP"""
|
||
# 按类别分组
|
||
categories = set([inst['label'] for inst in self.gt])
|
||
|
||
aps = []
|
||
for category in categories:
|
||
pred_cat = [p for p in self.pred if p['label'] == category]
|
||
gt_cat = [g for g in self.gt if g['label'] == category]
|
||
|
||
# 按置信度排序
|
||
pred_cat = sorted(pred_cat, key=lambda x: x['confidence'], reverse=True)
|
||
|
||
# 计算precision-recall
|
||
tp = np.zeros(len(pred_cat))
|
||
fp = np.zeros(len(pred_cat))
|
||
|
||
matched_gt = set()
|
||
|
||
for i, pred in enumerate(pred_cat):
|
||
best_iou = 0
|
||
best_gt_idx = -1
|
||
|
||
for j, gt in enumerate(gt_cat):
|
||
if j in matched_gt:
|
||
continue
|
||
|
||
iou = self.compute_iou_3d(pred['bbox'], gt['bbox'])
|
||
if iou > best_iou:
|
||
best_iou = iou
|
||
best_gt_idx = j
|
||
|
||
if best_iou >= iou_threshold:
|
||
tp[i] = 1
|
||
matched_gt.add(best_gt_idx)
|
||
else:
|
||
fp[i] = 1
|
||
|
||
# 累积
|
||
tp_cumsum = np.cumsum(tp)
|
||
fp_cumsum = np.cumsum(fp)
|
||
|
||
recalls = tp_cumsum / len(gt_cat)
|
||
precisions = tp_cumsum / (tp_cumsum + fp_cumsum + 1e-8)
|
||
|
||
# 计算AP（11点插值）
|
||
ap = self.compute_ap(recalls, precisions)
|
||
aps.append(ap)
|
||
|
||
return np.mean(aps)
|
||
|
||
def evaluate_scene_graph(self, pred_graph, gt_graph):
|
||
"""场景图评测"""
|
||
# 节点准确率
|
||
node_precision = len(set(pred_graph.nodes) & set(gt_graph.nodes)) / len(pred_graph.nodes)
|
||
node_recall = len(set(pred_graph.nodes) & set(gt_graph.nodes)) / len(gt_graph.nodes)
|
||
|
||
# 边准确率
|
||
edge_precision = len(set(pred_graph.edges) & set(gt_graph.edges)) / len(pred_graph.edges)
|
||
edge_recall = len(set(pred_graph.edges) & set(gt_graph.edges)) / len(gt_graph.edges)
|
||
|
||
return {
|
||
'node_precision': node_precision,
|
||
'node_recall': node_recall,
|
||
'edge_precision': edge_precision,
|
||
'edge_recall': edge_recall
|
||
}
|
||
```
|
||
|
||
### 8.3 物理交互评测
|
||
|
||
```python
|
||
class PhysicsEvaluator:
|
||
def __init__(self, simulator, world_model):
|
||
self.sim = simulator
|
||
self.model = world_model
|
||
|
||
def evaluate_prediction_accuracy(self, test_scenarios):
|
||
"""评估物理预测准确性"""
|
||
results = []
|
||
|
||
for scenario in test_scenarios:
|
||
# 初始状态
|
||
initial_state = scenario['initial_state']
|
||
action = scenario['action']
|
||
|
||
# 真实仿真结果
|
||
self.sim.set_state(initial_state)
|
||
self.sim.apply_action(action)
|
||
self.sim.step(n_steps=100)
|
||
true_final_state = self.sim.get_state()
|
||
|
||
# 世界模型预测
|
||
predicted_final_state = self.model.predict(
|
||
initial_state, action, n_steps=100
|
||
)
|
||
|
||
# 计算误差
|
||
position_error = np.linalg.norm(
|
||
true_final_state['positions'] - predicted_final_state['positions']
|
||
)
|
||
|
||
velocity_error = np.linalg.norm(
|
||
true_final_state['velocities'] - predicted_final_state['velocities']
|
||
)
|
||
|
||
results.append({
|
||
'scenario': scenario['name'],
|
||
'position_error': position_error,
|
||
'velocity_error': velocity_error
|
||
})
|
||
|
||
return results
|
||
|
||
def evaluate_robot_task_success(self, tasks):
|
||
"""评估机器人任务成功率"""
|
||
success_count = 0
|
||
|
||
for task in tasks:
|
||
# 执行任务
|
||
success = self.execute_task_with_model(task)
|
||
|
||
if success:
|
||
success_count += 1
|
||
|
||
success_rate = success_count / len(tasks)
|
||
|
||
return {
|
||
'success_rate': success_rate,
|
||
'total_tasks': len(tasks),
|
||
'successful_tasks': success_count
|
||
}
|
||
|
||
def evaluate_zero_shot_generalization(self, novel_scenarios):
|
||
"""评估零样本泛化能力"""
|
||
# 在未见过的物体/场景上测试
|
||
results = []
|
||
|
||
for scenario in novel_scenarios:
|
||
# 使用世界模型进行规划
|
||
plan = self.model.plan(scenario['goal'])
|
||
|
||
# 执行并评估
|
||
success = self.execute_plan(plan, scenario)
|
||
|
||
results.append({
|
||
'scenario': scenario['name'],
|
||
'success': success,
|
||
'plan_length': len(plan)
|
||
})
|
||
|
||
return results
|
||
```
|
||
|
||
### 8.4 综合评测基准
|
||
|
||
```yaml
|
||
评测套件: HotelScene-Bench
|
||
|
||
子基准1_几何重建:
|
||
数据集: 10个酒店场景（公共区域 + 客房 + 卫生间）
|
||
指标:
|
||
- Chamfer Distance < 3cm
|
||
- Accuracy@5cm > 95%
|
||
- Completeness@5cm > 90%
|
||
- Normal Consistency > 0.85
|
||
|
||
子基准2_渲染质量:
|
||
测试视角: 每场景100个新视角
|
||
指标:
|
||
- PSNR > 28 dB
|
||
- SSIM > 0.85
|
||
- LPIPS < 0.15
|
||
- FPS > 30 @ 1080p
|
||
|
||
子基准3_语义理解:
|
||
标注物体: 500+实例
|
||
指标:
|
||
- 3D mAP@0.5 > 70%
|
||
- 场景图节点F1 > 0.80
|
||
- 场景图边F1 > 0.65
|
||
- 开放词表检测准确率 > 60%
|
||
|
||
子基准4_物理交互:
|
||
任务类型:
|
||
- 导航（10个场景）
|
||
- 物体操作（50个任务）
|
||
- 铰接物体交互（30个任务）
|
||
指标:
|
||
- 导航成功率 > 85%
|
||
- 抓取成功率 > 75%
|
||
- 开门/拉抽屉成功率 > 80%
|
||
- 物理预测误差 < 10cm
|
||
|
||
子基准5_卫生间专项:
|
||
场景: 5个高反光卫生间
|
||
指标:
|
||
- 镜面几何恢复准确率 > 80%
|
||
- 玻璃透明物体检测率 > 70%
|
||
- 水龙头操作成功率 > 75%
|
||
```
|
||
|
||
---
|
||
|
||
## 九、项目实施时间规划
|
||
|
||
### 9.1 总体时间线（12个月）
|
||
|
||
```mermaid
|
||
gantt
|
||
title 酒店场景建模项目甘特图
|
||
dateFormat YYYY-MM-DD
|
||
|
||
section 准备阶段
|
||
硬件采购与到货 :p1, 2026-06-01, 30d
|
||
传感器标定与测试 :p2, after p1, 14d
|
||
软件环境搭建 :p3, 2026-06-01, 21d
|
||
|
||
section 数据采集
|
||
公共区域采集 :d1, after p2, 14d
|
||
客房采集（5间） :d2, after d1, 21d
|
||
卫生间采集 :d3, after d2, 14d
|
||
|
||
section 算法开发
|
||
SLAM建图模块 :a1, after p3, 30d
|
||
3DGS训练流程 :a2, after a1, 21d
|
||
语义理解模块 :a3, after a2, 30d
|
||
Ref-NeRF实现 :a4, after a3, 21d
|
||
|
||
section 系统集成
|
||
场景图构建 :i1, after a3, 21d
|
||
物理仿真集成 :i2, after i1, 30d
|
||
M-JEPA训练 :i3, after i2, 45d
|
||
|
||
section 验证与优化
|
||
评测基准构建 :v1, after i2, 21d
|
||
机器人实验 :v2, after i3, 30d
|
||
系统优化迭代 :v3, after v2, 30d
|
||
|
||
section 成果输出
|
||
论文撰写 :o1, after v2, 60d
|
||
开源准备 :o2, after v3, 21d
|
||
文档编写 :o3, after o2, 14d
|
||
```
|
||
|
||
### 9.2 详细里程碑
|
||
|
||
| 月份 | 里程碑 | 关键交付物 | 验收标准 |
|
||
|-----|--------|-----------|---------|
|
||
| **M1** | 项目启动 | 硬件到位、环境搭建完成 | 传感器标定误差 < 5mm |
|
||
| **M2** | 数据采集完成 | 3个场景原始数据 | 数据完整性 > 95% |
|
||
| **M3** | SLAM建图验证 | 公共区域点云地图 | 闭环误差 < 0.5% |
|
||
| **M4** | 3DGS训练完成 | 实时渲染模型 | PSNR > 26 dB |
|
||
| **M5** | 语义理解集成 | 场景图数据库 | mAP@0.5 > 60% |
|
||
| **M6** | 卫生间模块完成 | Ref-NeRF模型 | 镜面恢复准确率 > 75% |
|
||
| **M7** | 物理仿真就绪 | Isaac Sim场景 | 可交互物体 > 50个 |
|
||
| **M8** | M-JEPA训练完成 | 世界模型权重 | 预测误差 < 15cm |
|
||
| **M9** | 机器人实验 | 任务成功率报告 | 综合成功率 > 70% |
|
||
| **M10** | 系统优化 | 优化后模型 | 性能提升 > 20% |
|
||
| **M11** | 论文投稿 | 会议论文 | 投稿至顶会 |
|
||
| **M12** | 开源发布 | GitHub仓库、数据集 | 文档完整度 100% |
|
||
|
||
### 9.3 人力资源配置
|
||
|
||
```yaml
|
||
团队组成（建议）:
|
||
|
||
项目负责人（PI）× 1:
|
||
- 职责: 总体把控、对外合作、论文指导
|
||
- 投入: 20% 时间
|
||
|
||
算法工程师 × 2:
|
||
- 职责: SLAM、3DGS、NeRF算法实现与优化
|
||
- 技能: C++/Python、CUDA、计算机视觉
|
||
- 投入: 100% 时间
|
||
|
||
机器学习工程师 × 1:
|
||
- 职责: M-JEPA训练、语义理解模块
|
||
- 技能: PyTorch、Transformer、强化学习
|
||
- 投入: 100% 时间
|
||
|
||
机器人工程师 × 1:
|
||
- 职责: Isaac Sim集成、机器人实验
|
||
- 技能: ROS2、物理仿真、机械臂控制
|
||
- 投入: 100% 时间
|
||
|
||
数据工程师 × 1:
|
||
- 职责: 数据采集、标注、数据库管理
|
||
- 技能: 传感器操作、SQL、数据处理
|
||
- 投入: 80% 时间（前6个月100%）
|
||
|
||
研究助理 × 2:
|
||
- 职责: 实验辅助、评测、文档编写
|
||
- 投入: 50% 时间
|
||
```
|
||
|
||
---
|
||
|
||
## 十、风险管理与应对策略
|
||
|
||
### 10.1 技术风险
|
||
|
||
| 风险项 | 概率 | 影响 | 应对策略 |
|
||
|-------|------|------|---------|
|
||
| **卫生间高反光重建失败** | 中 | 高 | 1. 提前进行小规模测试<br>2. 准备备选方案（手动建模）<br>3. 与偏振相机厂商技术支持 |
|
||
| **3DGS训练不收敛** | 低 | 中 | 1. 使用成熟的开源实现<br>2. 调整学习率和初始化<br>3. 分块训练降低难度 |
|
||
| **M-JEPA物理预测不准确** | 中 | 中 | 1. 增加仿真数据量<br>2. 引入物理先验约束<br>3. 降低预测时长要求 |
|
||
| **实时性能不达标** | 低 | 中 | 1. 使用LOD技术<br>2. 模型剪枝与量化<br>3. 升级GPU硬件 |
|
||
| **语义分割精度低** | 低 | 低 | 1. 人工标注补充训练<br>2. 使用更大的基础模型<br>3. 多模型集成 |
|
||
|
||
### 10.2 工程风险
|
||
|
||
| 风险项 | 概率 | 影响 | 应对策略 |
|
||
|-------|------|------|---------|
|
||
| **硬件故障或延期** | 中 | 高 | 1. 提前2周下单<br>2. 准备备用设备<br>3. 建立供应商备选清单 |
|
||
| **数据采集权限受限** | 中 | 高 | 1. 提前与酒店沟通协议<br>2. 准备数据脱敏方案<br>3. 考虑使用公开数据集 |
|
||
| **存储空间不足** | 低 | 中 | 1. 实时监控存储使用<br>2. 及时清理中间文件<br>3. 扩展云存储 |
|
||
| **团队成员离职** | 低 | 高 | 1. 代码文档化<br>2. 知识定期分享<br>3. 关键模块双人备份 |
|
||
|
||
### 10.3 进度风险
|
||
|
||
| 风险项 | 概率 | 影响 | 应对策略 |
|
||
|-------|------|------|---------|
|
||
| **数据采集超时** | 中 | 中 | 1. 预留20%缓冲时间<br>2. 并行采集多个场景<br>3. 简化采集流程 |
|
||
| **算法调试周期长** | 高 | 中 | 1. 使用小规模数据快速迭代<br>2. 设置阶段性目标<br>3. 及时调整技术路线 |
|
||
| **论文被拒需返工** | 中 | 低 | 1. 提前内部审稿<br>2. 准备多个投稿目标<br>3. 持续改进实验 |
|
||
|
||
---
|
||
|
||
## 十一、预算估算
|
||
|
||
### 11.1 硬件设备预算
|
||
|
||
| 类别 | 项目 | 数量 | 单价（万元） | 小计（万元） |
|
||
|-----|------|------|------------|------------|
|
||
| **传感器** | Livox Mid-360 LiDAR | 1 | 1.5 | 1.5 |
|
||
| | Azure Kinect DK | 2 | 0.3 | 0.6 |
|
||
| | 偏振相机 Lucid Phoenix | 1 | 3.0 | 3.0 |
|
||
| | 高分辨率相机 Sony α7R IV | 1 | 2.0 | 2.0 |
|
||
| | IMU Xsens MTi-630 | 1 | 1.2 | 1.2 |
|
||
| **平台** | 移动采集小车（定制） | 1 | 2.0 | 2.0 |
|
||
| | Matterport Pro3（可选） | 1 | 4.0 | 0 |
|
||
| **计算** | 工作站（2×RTX 4090） | 2 | 4.0 | 8.0 |
|
||
| | NVIDIA Jetson AGX Orin | 1 | 1.0 | 1.0 |
|
||
| |