Pocket Naturalist – Offline AI field guide for walks

说明与用途

Pocket Naturalist 是一款离线植物识别指南，使用 Gemma 4 12B 模型。添加 107 条印度植物的本地字段指南后，准确率从 4/15 提升至 10/15，并消除了虚假名称和伪造事实。

问题

Users need an offline plant identification guide that works without internet, provides reliable results, and encourages real‑world observation.

结果

Adding a region‑specific field guide CSV and prompt rules increased correct plant identifications from 4/15 to 10/15 and eliminated fabricated facts, while keeping the system fully offline.

工作流

• 1. Start a walk – the phone receives a first quest via offline TTS.
• 2. Snap a photo or add from gallery; the phone connects to the laptop over a hotspot.
• 3. The laptop runs server.py, which sends the image to Ollama running gemma4:12b (vision).
• 4. The model outputs a fixed‑JSON response: features seen, top three candidates, and a "look‑closer" clue if unsure.
• 5. The app reads the answer aloud, then returns the phone to the pocket.
• 6. At walk end, the app compiles a journal from the collected data and user corrections.

证据

Stock Gemma got 4 of 15 right.

来源

After the fix: 4 → 10 correct out of 15 photos where I know the right answer.

来源

The laptop is nothing special: an RTX 3050 with 6 GB of GPU memory and 16 GB of RAM.

来源