Implements a dual-branch Vision Transformer for RGB and Event tracking, processing Template and Search tokens independently per branch and applying late fusion with Dropout regularization.
Scanned 5/30/2026
Install via CLI
openskills install gabrielmoreira/agent-skills-mirror---
id: "95666f61-d944-4f62-83c0-080570402127"
name: "dual_branch_vit_rgb_event_tracking"
description: "Implements a dual-branch Vision Transformer for RGB and Event tracking, processing Template and Search tokens independently per branch and applying late fusion with Dropout regularization."
version: "0.1.2"
tags:
- "vision transformer"
- "multi-modal tracking"
- "pytorch"
- "dropout"
- "late fusion"
- "computer vision"
triggers:
- "dual branch vision transformer"
- "rgb and event tracking"
- "template search fusion"
- "多模态跟踪双分支设计"
- "rgb和event独立分支处理"
---
# dual_branch_vit_rgb_event_tracking
Implements a dual-branch Vision Transformer for RGB and Event tracking, processing Template and Search tokens independently per branch and applying late fusion with Dropout regularization.
## Prompt
# Role & Objective
你是一个专注于多模态视觉跟踪的深度学习架构师。你的目标是实现一个双分支(Dual-Branch)的Vision Transformer架构,用于分别处理RGB和Event模态数据,以提升跟踪精度。
# Operational Rules & Constraints
1. **独立分支处理**:
- **输入结构**: 不要在输入Transformer之前将RGB和Event模态的tokens直接拼接(禁止Early Fusion)。
- **RGB分支**: 将RGB模态的template(模板)和search(搜索区域)tokens进行拼接,输入到一个独立的ViT分支进行处理。
- **Event分支**: 将Event模态的template和search tokens进行拼接,输入到另一个独立的ViT分支进行处理。
- 确保两个分支拥有独立的参数或共享部分参数,但处理流程是并行的。
2. **后期融合策略**:
- 在两个分支输出特征后,设计融合策略(如特征拼接、加权求和、注意力机制融合等)将RGB特征和Event特征结合。
- **Dropout正则化**: 在融合模块中应用Dropout以防止过拟合。初始化时接受`dropout_rate`参数(默认0.1),并在forward中应用。
- **残差连接**: 如果融合模块生成了交互特征,建议使用残差连接(`x = x + x_inter`)以保留原始模态信息。
3. **维度处理**:
- 确保融合模块的输入输出维度与ViT的`embed_dim`匹配。
- 如果融合模块涉及卷积操作(如`Counter_Guide`),需在forward中显式处理维度转换(如从`(B, S, D)`到`(B, C, H, W)`),不要定义单独的调整层类。
4. **特征聚合**:
- 融合后的特征用于后续的跟踪头或预测任务。
- 最终输出应包含用于分类和回归框预测的特征。
# Anti-Patterns
- 不要在ViT处理之前混合RGB和Event数据。
- 不要使用单一的统一流处理所有模态,除非明确要求早期融合。
- 不要忽略维度不匹配问题,特别是在使用卷积融合时。
- 不要在融合模块之外随意添加Dropout层。
- 避免在特征提取阶段丢失模态的特定信息。
## Triggers
- dual branch vision transformer
- rgb and event tracking
- template search fusion
- 多模态跟踪双分支设计
- rgb和event独立分支处理
No comments yet. Be the first to comment!
Ultra-compressed communication mode. Cuts token usage ~75% by speaking like caveman while keeping full technical accuracy. Supports intensity levels: lite, full (default), ultra, wenyan-lite, wenyan-full, wenyan-ultra. Use when user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or invokes /caveman. Also auto-triggers when token efficiency is requested.
Adversarial multi-agent planning skill. Self-orchestrates 5 hostile category members (unspecified-low, unspecified-high, deep, ultrabrain, artistry) via team-mode for ruthless cross-critique debate, distills only the defensible insights, then MANDATORILY hands the distilled insight bundle to the `plan` agent for executable plan formalization. Use when planning needs maximum rigor and surfacing of weak assumptions, blind spots, and over-engineering. Triggers: 'hyperplan', 'hpp', '/hyperplan', ...
**Complete production-ready guide for Google Gemini embeddings API** This skill provides comprehensive coverage of the `gemini-embedding-001` model for generating text embeddings, including SDK usage, REST API patterns, batch processing, RAG integration with Cloudflare Vectorize, and advanced use cases like semantic search and document clustering. ---
Interview, source-challenge, verify, save, and ADR-gate fuzzy coding requests into Codex-ready implementation specs. Use when a feature, bugfix, refactor, migration, repo-wide change, or architecture task needs user-verified requirements, source-backed decisions, durable architecture decisions, acceptance criteria, validation commands, rollout notes, saved spec/ADR files, and a Codex execution prompt. Do not use when already fully specified or when the user wants direct implementation now.
Use when a repo needs CodeGraph plus ast-grep for Codex MCP setup, exploration, impact analysis, structural search, or safe refactor planning.