ORTUS AI开源RightWayUp：360度图像旋转检测模型及JPEG基
Open-sourcing RightWayUp - a 360-degree image rotation model, and a JPEG shortcut we found in a common benchmark [P]
不仅开源了实用的360度旋转检测模型，更揭示了常见基准测试中的JPEG压缩漏洞，对评估图像模型鲁棒性有重要参考价值，建议相关研究者关注。
We are open-sourcing RightWayUp - a new image rotation detection model.
我们开源了 RightWayUp——一种新的图像旋转检测模型。
I work at ORTUS AI (we develop video analytics). We needed to tell from a single CCTV frame whether a camera had been rotated or installed at an angle (or upside-down), so we tried different models for that, without much luck (low accuracy, lots of false positives on regular camera-like frames). Some were not permissively licensed. Out of despair, we decided to train our own. We didn't expect it to come out this good.
我在 ORTUS AI 工作（我们开发视频分析技术）。我们需要从单个 CCTV 帧中判断摄像头是否被旋转、倾斜安装或倒置，因此我们尝试了不同的模型，但效果不佳（准确率较低，在常规摄像头画面中误报较多）。其中一些模型的许可证不够宽松。出于无奈，我们决定自行训练一个模型。没想到效果会这么好。
It is now called RightWayUp. It estimates how far an image is rotated from upright, all 360°, and abstains when there's no clear "up" (sky, ground, close-ups, etc.). We are releasing code and weights under Apache-2.0, in six sizes from Pico (small enough to run in a browser) to Max (ahead of the other models we tried on most of our tests).
现在它被称为 RightWayUp。它能估算图像相对于正立状态的旋转角度，覆盖全部 360°；当没有明确的“上方”指示时（如天空、地面、特写等），它会选择不做预测。我们以 Apache-2.0 许可证发布代码和权重，提供六种尺寸，从 Pico（小到足以在浏览器中运行）到 Max（在我们测试的大多数场景中优于其他我们尝试过的模型）。
We tested it on many test images, including held-out ones (which the model never saw before). In tests with the held-out set, RightWayUp Max was within 10° on 93.0% of them vs 88.4% for Woehrer 2026 (one of the better other models we liked). In the Woehrer 2026 own COCO-based benchmark, it was 98.8% accurate within 10° (five-seed mean) vs Woehrer's 98.0%. It also gets every image right on RotBench.
我们在许多测试图像上对其进行了测试，包括预留的测试集（模型此前从未见过）。在预留测试集的测试中，RightWayUp Max 有 93.0% 的图像误差在 10° 以内，而 Woehrer 2026（我们喜欢的其他较好模型之一）为 88.4%。在 Woehrer 2026 基于 COCO 的基准测试中，其误差在 10° 以内的准确率为 98.8%（五种子种子均值），而 Woehrer 为 98.0%。它在 RotBench 上也实现了全对。
While training and testing, we found something we didn't expect: saving the images from the COCO-based rotation benchmark as JPEG q90 drops Woehrer 2026 from 98.0% to 30.2% (five-seed mean), while our models barely move. We suspect the rotated JPEG grid of the source photos gives the angle away. We found that in our early experiments, and made an effort to remove its effect.
在训练和测试过程中，我们发现了一个意想不到的现象：将基于 COCO 的旋转基准测试中的图像保存为 JPEG q90 格式后，Woehrer 2026 的准确率从 98.0% 骤降至 30.2%（五种子种子均值），而我们的模型几乎不受影响。我们怀疑源照片的旋转 JPEG 网格泄露了角度信息。我们在早期实验中发现了这一点，并努力消除了其影响。
I hope it will be useful to the community.
希望它对社区有用。
For transparency: parts of the engineering were done with Claude and Codex.
为透明起见：部分工程工作由 Claude 和 Codex 协助完成。
Here is a full write-up with a demo video: https://cheqit.ortusai.io/resources/rightwayup/
完整文章及演示视频见：https://cheqit.ortusai.io/resources/rightwayup/
Code (with a link to the weights): https://github.com/ortusaitech/rightwayup
代码（含权重链接）：https://github.com/ortusaitech/rightwayup
Questions and failure cases welcome.
欢迎提问和反馈失败案例。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力