HackerNews中文版

你好，HN，我从事结构检查的视觉系统工作。一个常见的问题是，虽然我们有很多“健康”的图像，但通常缺乏一个可靠的“黄金集”，其中包含稀有故障（如破碎的瓷器），以便在部署之前验证我们的模型。例如，如果你的测试集仅包含5个故障模式的样本，你就无法信任模型的召回率。为了解决这个问题，我建立了一个生成数据集的管道。在这个例子中，我使用了7个真实的缺陷样本，提取了它们的拓扑/纹理，并程序性地生成了200个在不同光照和背景下难以检测的变体。我发布这一批破损绝缘子（CC0），特别是为了帮助团队基准测试他们模型在稀有类别上的召回率： - 输入：7个真实样本。 - 输出：200个完全标注的评估图像（COCO/YOLO）。 - 用例：验证/测试集（不是完整的训练集）。你们目前是如何验证“1万分之一”边缘案例的召回率的？ Jérôme

查看原文

Hello HN,I work on vision systems for structural inspection. A common pain point is usually that while we have a lot of "healthy" images, we often lack a reliable "Golden Set" of rare failures (like shattered porcelain) to validate our models before deployment.You can't trust your model's recall if your test set only has 5 examples of the failure mode for example.So to fix this, I built a pipeline to generate datasets. In this example, I took 7 real-world defect samples, extracted their topology/texture, and procedurally generated 200 hard-to-detect variations across different lighting and backgrounds.I’m releasing this batch of broken insulators (CC0) specifically to help teams benchmark their model's recall on rare classes:<a href="https://www.silera.ai/blog/free-200-broken-insulators-dataset" rel="nofollow">https://www.silera.ai/blog/free-200-broken-insulators-datase...</a>- Input: 7 real samples.- Output: 200 fully labeled evaluation images (COCO/YOLO).- Use Case: Validation / Test Set (not full training).How do you guys currently validate recall for "1 in 10,000" edge cases?Jérôme

展示HN：我从7张真实照片中生成了200个罕见缺陷的“压力测试”