实测双曲嵌入对图像搜索排序的影响：Hyper3-CLIP V1 对比 OpenAI CLIP
Milvus 拿 5 万张图实测了双曲嵌入和 CLIP 的检索差距，写描述找图它更强，搜大类还是 cosine 稳，动手前可以看看。
Milvus 团队用 5 万张 ImageNet 验证图测试 Hyper3-CLIP V1：Lorentz 距离比 cosine 将分类准确率从 50.14% 提到 51.06%，但在 132 个大类搜索上落后 cosine 2.2 平均精度点。与 OpenAI CLIP 相比，V1 在描述性标题检索和区分“狗在沙发上/沙发在狗上”更准，CLIP 则在大类搜索领先 6.4 平均精度点，差距主要来自稀有分类词。较小的 COCO 测试覆盖三个概念时 V1 反超。结论是描述性标题搜索可试用 V1，大类搜索建议在自己的数据上对比 Lorentz 与 cosine。
We tested how hyperbolic embeddings change image search rankings. First, we compared Hyper3-CLIP V1 with OpenAI CLIP. Then we scored the same V1 embeddings with Lorentz (full hyperbolic geometry) and cosine (direction only) to isolate the effect of the geometry. On 50,000 ImageNet validation images, Lorentz raised V1’s classification accuracy from 𝟱𝟬.𝟭𝟰% 𝘁𝗼 𝟱𝟭.𝟬𝟲%. But across 132 broad category searches such as “bird,” it trailed cosine by 𝟮.𝟮 𝗮𝘃𝗲𝗿𝗮𝗴𝗲 𝗽𝗿𝗲𝗰𝗶𝘀𝗶𝗼𝗻 𝗽𝗼𝗶𝗻𝘁𝘀. Its remaining classification errors did not land closer to the right category either. Against CLIP, V1 did better at finding the right image for a descriptive caption and at distinguishing “a dog on a sofa” from “a sofa on a dog.” CLIP led by 6.4 average precision points on our ImageNet broad category searches, with rare taxonomy terms accounting for much of the gap. In a smaller COCO test covering three broad concepts, V1 led instead. Our takeaway: Hyper3-CLIP V1 is worth testing when users search for images with descriptive captions. For broad category search, evaluate the words your users actually type and compare Lorentz with cosine on your own data. Milvus can run hyperbolic ranking with a standard inner product milvus.io/blog/what-hype… xperiment: https://t.co/UWlkSaC3Tw 💬 0 🔄 0 ❤️ 0 👀 190 ⚡
