Grassdex: Building a 100% Offline Nature Dex & Touch-Grass Bingo with Local Gemma

说明与用途

Grassdex 是一款离线自然识别应用，使用本地 Gemma 4 通过 Ollama 进行推理，识别物种并生成 3×3 的“Touch‑Grass Bingo”卡片，确保 45 分钟徒步中屏幕时间不超过 4 分钟。

问题

Outdoor nature identification apps rely on cloud servers, causing long wait times, privacy concerns, and loss of service in low‑signal areas.

结果

Grassdex achieved 65 % strict accuracy and 94 % lenient accuracy on 17 test snaps, with 0 % JSON parse failures and median inference latency of 16.1 s. The app kept screen time under 4 min during a 45‑min walk.

工作流

• 1. Phone camera captures image and sends it over local Wi‑Fi to the laptop.
• 2. Flask app forwards image to Ollama running Gemma 4.
• 3. Single model call returns JSON with species name, confidence, facts, and bingo verification.
• 4. UI displays result and updates bingo card; when a line is completed, it shows "Bingo! Now put the phone away."

证据

Grassdex is a zero‑cloud local web app where your phone camera snaps nature over a private hotspot to your laptop, where an open‑weight vision model (Gemma 4 via Ollama) names the species and checks off a model‑generated 3x3 touch‑grass bingo card.

来源

Total screen‑on time was under 4 minutes across the entire 45‑minute walk (< 9% of total time).

来源

Strict accuracy (exact species/subject): 65% (11/17, Wilson 95% CI: 41% – 83%).

来源

Gemma’s raw confidence scores are uncalibrated… When Gemma 4 was dead wrong, it was just as confident as when it was right.

来源