File size: 7,353 Bytes
3fe13e0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
# TopoSlots Text Prompt 设计方案

## 核心问题

我们的模型接收 **text + target_skeleton** 双条件输入。text prompt 该怎么设计?

```
生成管线: text prompt + skeleton graph → SlotAssignment → MaskGIT → Decoder → motion
                 ↑                          ↑
            描述什么动作                  决定什么骨架
```

## 设计决策

### 决策 1:拓扑信息是否放入 prompt?

**结论:不放**。理由:

1. **架构已分离**:骨架拓扑通过 skeleton graph 独立输入(SkeletonEncoder + SlotAssignment),text 不需要重复编码拓扑
2. **跨骨架泛化**:如果 prompt 包含 "a dog walks",模型就被绑定到 dog 骨架。但 "walks forward" 可以复用到任何四足/双足
3. **可用性**:用户在推理时可以自由组合 "任意 text × 任意 skeleton",例如把 "walks forward" 应用到新见的骨架上
4. **标注效率**:不需要为每个骨架重写 prompt

**但有一个例外**:物种名可以作为**可选前缀标签**出现,用于训练时的语义对齐,推理时可以省略。

```
训练时:  "[dog] walks forward and sits down"     ← 方括号内为可选物种标签
推理时:  "walks forward and sits down"           ← 纯动作描述
推理时:  "[cat] walks forward and sits down"     ← 也可以加物种提示
```

### 决策 2:动作描述粒度

三个级别,**只要求 L1,鼓励 L2,L3 可选**:

| 级别 | 格式 | 示例 | 标注成本 | 用途 |
|------|------|------|:--------:|------|
| **L1: 动作标签** | 1-3 个词 | `walk`, `run`, `attack`, `idle` | 极低 | 分类、检索、baseline 条件 |
| **L2: 短描述** | 1 句话,10-20 词 | `walks forward slowly then stops` | 中等 | 主要训练条件 |
| **L3: 详细描述** | 2-3 句话 | `The creature begins walking at a steady pace, gradually slowing down. It pauses briefly, shifts weight, then stops.` | 高 | 精细控制、评估 |

**L2 是核心**——在标注成本和语义丰富度之间取得平衡。

### 决策 3:描述应该是骨架无关的(Skeleton-Agnostic)

**关键原则:描述"做什么"而不是"怎么做"。**

| ✅ 好的描述 (骨架无关) | ❌ 差的描述 (骨架特定) |
|---|---|
| `walks forward` | `moves left leg then right leg alternately` |
| `attacks with its mouth` | `opens jaw joint 15 degrees then closes` |
| `flies in a circle` | `flaps left wing up 45 degrees, right wing follows` |
| `stands idle, looking around` | `rotates head joint on spine3` |
| `runs and jumps over obstacle` | `extends both hind legs while fore legs tuck` |

**为什么**:
- "walks forward" 对人类 22 关节、狗 55 关节、蜘蛛 71 关节都有意义
- "moves left leg then right leg" 假设了双足结构,蜘蛛 8 条腿怎么办?
- 模型的 slot assignment + decoder 会自己决定哪些关节参与 "walk"

### 决策 4:物种通用动作词汇表

标准化动作词汇,跨物种通用:

```
== 运动类 ==
walk, run, sprint, trot, gallop, crawl, slither, swim, fly, hover, glide

== 姿态类 ==
stand, sit, lie down, crouch, squat, rear up, perch

== 交互类 ==
attack, bite, claw, kick, charge, ram, pounce, grab

== 情绪/状态类 ==
idle, alert, sleep, eat, drink, shake, scratch, groom

== 转场类 ==
turn left/right, start, stop, accelerate, decelerate, transition
```

这些词对人类和动物都有意义(人类也可以 "crouch", "charge", "idle")。

---

## 标注规范 (更新版)

### 每条 motion 的标注格式

```json
{
  "motion_id": "Dog_0001",
  "labels": {
    "L1_action": "walk",
    "L1_action_secondary": "turn",
    "L2_short": "walks forward then turns right",
    "L3_detailed": "The creature walks forward at a moderate pace for several steps, then smoothly turns to the right while maintaining balance.",
    "species_tag": "dog",
    "species_category": "quadruped"
  }
}
```

### 标注优先级

| 优先级 | 数据集 | 当前状态 | 需要做什么 |
|:------:|--------|---------|-----------|
| **P0** | HumanML3D | 100% L2 文本 | 补充 L1 标签(可从文本自动提取) |
| **P1** | Truebones Zoo | 80% L2 文本 | 补全剩余 20%;补充 L1 标签 |
| **P2** | LAFAN1 | 0% | 从文件名提取 L1(`aiming`, `dance`, `fight`...);VLM 生成 L2 |
| **P2** | 100Style | 0% | 从文件名提取 L1 + style 标签(`happy_walk`, `tired_run`...) |
| **P3** | Bandai Namco | 0% | 有 JSON metadata 可提取;VLM 补 L2 |
| **P3** | CMU MoCap | 0% | 从目录名提取 L1(subject/action 编号) |
| **P4** | Mixamo | 0% | 从文件名 hash→原始动画名映射 |

### 低成本批量标注方案

**Phase 1:自动提取 L1 标签**(零人工成本)
```
LAFAN1 文件名: "aiming1_subject1.bvh" → L1="aim"
100Style:      "Happy_FW.bvh"         → L1="walk", style="happy"
Bandai Namco:  JSON metadata           → L1 直接可用
CMU:           目录结构                 → L1 分类
```

**Phase 2:VLM 自动生成 L2 短描述**
- 渲染骨架动画为 stick figure 视频
- 送入 VLM (GPT-4o / Qwen2.5-VL) 生成 L2 描述
- 模板 prompt: `"Describe the motion of this creature in one sentence. Focus on WHAT it does, not HOW its body parts move."`

**Phase 3:人工审核 + 修正**(只审核 VLM 输出,不从零写)

---

## 训练时的 Text Conditioning 策略

```python
# 训练时随机选择条件级别(鼓励模型学习多粒度)
if has_L3 and random() < 0.2:
    text = sample['L3_detailed']
elif has_L2 and random() < 0.7:
    text = sample['L2_short']
elif has_L1:
    text = sample['L1_action']
else:
    text = ""  # unconditional (CFG dropout)

# 可选:加物种标签前缀
if random() < 0.5 and has_species:
    text = f"[{sample['species_tag']}] {text}"
```

### Classifier-Free Guidance 设计

```
10% 概率: text = "" (完全无条件 → 学习 unconditional motion prior)
10% 概率: text = L1 only (极简条件 → 学习粗粒度控制)
60% 概率: text = L2 short (主要条件)
20% 概率: text = L3 detailed (精细条件)

skeleton 始终提供 (不 dropout skeleton 条件)
```

---

## 与竞争方法的对比

| 方法 | Text 设计 | 骨架条件 | 跨骨架泛化 |
|------|---------|---------|:---------:|
| **HumanML3D** | 自由文本,包含 "a person" | 无(固定 SMPL) | ✗ |
| **NECromancer** | VLM 自动文本 | skeleton 分离 | ✓ (tokenizer) |
| **T2M4LVO** | 多粒度文本,含物种名 | 无 | ✗ |
| **AnyTop** | 无文本条件 | skeleton 分离 | ✓ |
| **TopoSlots (ours)** | **L1+L2+L3 多粒度,骨架无关,可选物种标签** | **skeleton 分离** | **✓ (生成)** |

**我们的优势**1. 文本和骨架完全解耦 → "walk" 可以应用到任何骨架
2. 多粒度条件 → 支持从粗到细的控制
3. 可选物种标签 → 训练时提供弱监督语义,推理时灵活组合

---

## 总结:标注工作量评估

| 阶段 | 工作量 | 产出 |
|------|--------|------|
| L1 自动提取 | ~1 天(脚本) | 所有数据集 100% L1 标签 |
| L2 VLM 生成 | ~2 天(GPU) | 缺失数据集 ~9K 条 L2 描述 |
| L2 人工审核 | ~3-5 天 | 质量保证 |
| L3 详细描述 | 可选 | 仅 HumanML3D 已有 |

**结论:L1 可以零成本自动化,L2 用 VLM 生成+人工审核,L3 暂不需要。总人工投入约 3-5 天。**