BEAST3D: Self-supervised 3D animal behavioral analysis and neural encoding from multi-view video via Gaussian splatting. Vision transformer predicts 3D Gaussian splats for differentiable rendering with simultaneous animal-background segmentation. Works with as few as 4 calibrated views.
Scanned 9/11/2026
Install to Claude Code
npx -y skills add hiyenwong/ai_collection --skill beast3d-gaussian-splatting-behavior --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Beast3d Gaussian Splatting Behavior?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/hiyenwong-beast3d-gaussian-splatting-behavior-ai-collection)More formats (shields.io, HTML) on the badges page.
---
name: beast3d-gaussian-splatting-behavior
description: "BEAST3D: Self-supervised 3D animal behavioral analysis and neural encoding from multi-view video via Gaussian splatting. Vision transformer predicts 3D Gaussian splats for differentiable rendering with simultaneous animal-background segmentation. Works with as few as 4 calibrated views."
tags: gaussian-splatting, animal-behavior, 3D-reconstruction, self-supervised, multi-view-video, neural-encoding, behavioral-analysis, vision-transformer
arxiv_id: 2606.02937
authors: Yanchen Wang, Lenny Aharon, Wangshu Zhu, Kyle Daruwalla, Linghua Zhang, Jiaru Zou, Selmaan Chettih, Helen Hou, Liam Paninski, Matthew R Whiteway
date: 2026-06-01
keywords: BEAST3D, Gaussian splatting, multi-view video, 3D behavioral analysis, self-supervised learning, neural encoding, animal pose estimation, view synthesis, laboratory experiments
score: 8
---
# BEAST3D: Self-Supervised 3D Animal Behavioral Analysis via Gaussian Splatting
**arXiv: 2606.02937** (June 2026)
**Authors**: Yanchen Wang et al. (Paninski & Whiteway groups)
## 核心问题
**背景**: 多视角视频录制越来越用于捕获动物3D运动,但提取丰富3D表征仍面临挑战:
- 监督姿态估计需要大量人工标注
- 通用3D重建模型在实验室稀疏视角设置失败
- 现有方法无法处理专业实验场景
**关键挑战**:
1. 如何从无标注多视角视频学习3D表征?
2. 如何在稀疏视角(4-6视图)下重建3D结构?
3. 如何将3D特征用于神经编码分析?
## 核心发现
### 1. 稀疏视角3D重建能力
实验设置:4个校准视角(vs通用方法需要密集重叠视角)
**关键创新**: 直接利用已知相机参数(而非估计相机几何)
### 2. 自监督学习框架
**BEAST3D架构**:
- Vision Transformer encoder
- Predict 3D Gaussian splats (position, covariance, opacity, color)
- Differentiable rendering reconstructs held-out views
- Simultaneous animal-background segmentation
### 3. 下游任务迁移
三个下游任务验证:
- 新视角合成:验证学习表征质量
- 多视角姿态估计:稀疏关键点轨迹
- 神经编码:3D特征→神经活动映射
**跨物种验证**: 4种动物
## 技术架构
### 3D Gaussian Splatting表示
每个splat包含:
- μ: 3D位置
- Σ: 3×3协方差矩阵(椭球形状)
- α: 不透明度(分割信号)
- c: 颜色/特征向量
优势:
- 显式3D表征(vs隐式神经场)
- 可微分渲染(高效训练)
- 稀疏表示(内存效率)
### Vision Transformer编码器
关键设计:
- 直接条件化已知相机参数
- 无需从密集视角估计几何
- 适配实验室稀疏视角设置
## 应用场景
### 触发词
使用此技能时:
- 多视角动物行为分析
- 3D姿态估计无标注
- 稀疏视角3D重建
- 行为-神经编码映射
- Gaussian splatting行为建模
- 自监督行为表征学习
### 应用领域
1. 神经行为实验:多视角设置 → 3D行为提取
2. 神经编码研究:3D特征 → 神经活动预测
3. 行为分析工具:无标注姿态轨迹生成
4. 动物运动研究:跨物种3D行为建模
### 实验配置建议
**相机设置**:
- 最少4个视角(空间分布)
- 校准精度:±1mm位置误差
- 同步采集(时间戳对齐)
**Gaussian Splat参数**:
- Splat数量: 1000-5000(动物大小)
- 协方差初始化: 球形→迭代优化为椭球
- 不透明度阈值: 0.5(分割边界)
## 参考文献
- arXiv:2606.02937 (2026) - Wang et al.
- Kerbl et al. (2023) - 3D Gaussian Splatting
- Dosovitskiy et al. (2020) - Vision Transformer
---
**Activation**: gaussian-splatting, animal-behavior, 3D-reconstruction, self-supervised, multi-view-video, neural-encoding, behavioral-analysis, vision-transformerIs this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!