| # 🚀 Oculus V3: Future Training Roadmap | |
| COCO (Current) = 80 common classes. Good baseline, but limited for real-world niche tasks. | |
| ## Option A: Universal Detection (The "Scanner") | |
| **Target**: Detect 1200+ specific objects. | |
| - **Dataset**: **LVIS** or **Objects365**. | |
| - **Result**: Recognizes "stapler", "doorknob", "mango" instead of just generic classes. | |
| ## Option B: Visual Reasoning (The "Thinker") | |
| **Target**: Better VQA and complex instruction following. | |
| - **Dataset**: **LLaVA-Instruct** or **VizWiz**. | |
| - **Why**: Teaches the model to "explain why the car is parked" rather than just finding the car. | |
| - **Result**: A smarter chatbot-like VLM. | |
| ## Recommendation | |
| Since Oceanir is a VLM platform, **Option B (Instruction Tuning)** is the highest value next step. It improves the model's IO (Intelligence Output) significantly more than just adding more bounding boxes. | |