OpenBMB· @OpenBMB · X·· 2026-09-20AI 评分39
AI 导读
面壁智能 OpenBMB 的本地模型 Augury 经调优后,植物照片识别 top-1 准确率从 71.8% 提升至 80.2%,通过合并重复物种键和加入 PCA 白化实现。该模型还修复了格式化器编造 pH 值的问题,将置信度阈值从 0.80 下调至 0.65,澳大利亚带土壤数据的物种从 188 增至 221。下一步是开发 GUI 封装,让农民能在手机上直接使用。
正文 · 原文
This is the kind of local AI use case I love seeing.
Not just “AI can identify a plant,” but actually helping farmers make better decisions from a photo — entirely on-device.
80.2% → 90%+ top-1 is a big jump, but getting this into a simple phone GUI for farmers is where it gets really interesting 🌱
Local model - Augury, that genuinely help farmers make better farming decisions: Some evening tweaks on Augury @OpenBMB model. Here are the highlights, I believe we will reach 90%+ plant id accuracy by the end of the weekend. - Photo ID went from 71.8% to 80.2% top-1 by merging duplicate species keys and adding PCA whitening. - The formatter invented pH numbers in 7 of 20 answers; it now invents none because a structural pass catches and neutralises them. - The old 0.80 confidence threshold was backwards — precision peaks at 0.65 and falls above 0.75. - Australian species with soil data went from 188 to 221, including Paterson's curse and serrated tussock. - The extractor now resolves all 11 names that previously failed, via 158 curated common names. - Management questions like "what herbicide for lantana?" now decline instead of answering off-topic. - Two claims in my earlier report were wrong and are now corrected with visible notices. - 90% top-1 was not reached, and the measured reason is the encoder on CPU, not the method. We keep on going! Weeds as indicators model is improving. Next step - create a GUI wrapper so it can be shipped to those who will use it - farmers on their phone.在 X 查看被引用的帖子
来源:OpenBMB · x.com