将Oxford-Face Api与Face ++ Api和Open CV进行比较 [英] Comparing Project Oxford- Face Api with Face++ Api and Open CV

问题描述

我一直在尝试Project Oxford的Face Api进行面部检测和面部识别。项目牛津中的人脸识别效果很好，但问题出现在人脸检测部分。当用户站在离相机超过3英尺的地方时，脸部检测无法工作。为了演示，请考虑附加的两个样本图像。 [1] [2]

图像分辨率为320 X 240，它是灰度图像。在这两个图像中，受试者站在距离相机约3.5英尺处。在这两种情况下，面部api都无法检测到面部。 （您可以通过在此处发布图像来测试它[3]） 

与上面的图像一样，脸部++ api似乎效果很好。（您可以通过发布来测试它）图片  face ++ website  [4]）。在这两个图像中，面部的置信度超过99％。即使它检测到玻璃的价格超过95美元，也不会超过95％。 

使用opencv python库，面部检测距离相机的距离超过3英尺。当然，我没有opencv结果的置信度数据。

我已经测试过多次迭代，不同的主题站在摄像头前面。在所有情况下，项目牛津大学的脸api无法检测到脸部，脸部++和opencv成功。

我检查了面部api文档，看看是否有任何参数，如比例因子提高准确性。但是没有任何东西可用。

这是否得出结论项目oxford api中的面部检测准确性低于face ++和opencv，或者我在此比较中是否遗漏了什么？ 

[1]  https://social.msdn.microsoft.com/Forums/getfile/867475

[2]  https://social.msdn.microsoft.com/Forums/getfile/867476

[3]  https://www.microsoft.com/认知服务/ en-us / face-api

[4]  http：//www.faceplusplus.com/demo-detect/

PS：很抱歉单独发布链接。我的帐户尚未经过验证。 

解决方案

  最可能的原因是面部比例小于36x36， Face-Detection API中最小的可检测面部比例，请参考

https：// dev.projectoxford.ai/docs/services/563879b61984550e40cbbe8d/operations/563879b61984550f30395236 了解更多详情。

In为了检测面部，您可以在调用此API之前放大图像。

感谢您使用我们的服务。

Hi,

I have been trying out Project Oxford's Face Api for face detection and face identification. Face identification in project oxford works great, but the problem comes in the face detection part. When the user is standing more than 3 feet away from the camera, the face detection fails to work. In order to demonstrate, consider the two sample images attached. [1] [2]

The image resolution is 320 X 240 and it's a gray scale image. In these two images, the subject is standing at around 3.5 feet from the camera. In both the cases, the face api fails to detect the faces. (You can test it by posting the images here [3]) 

Where as with the above images, face++ api seems to works well. (You can test it by posting the images face ++ website [4]). In both the images the confidence of the face is more than 99%. Even it detects the glass with more than 95% confidence. 

With the opencv python library, the face detections works great at more than 3 feet distance from the camera. Of course, I donot have confidence data for opencv results.

I have tested with many iterations with different subjects standing infront of the camera. In all the cases, the project oxford's face api fails to detect the face, where as face++ and opencv succeeds.

I checked the face api document to see if there's any parameters like scale factors to improve the accuracy. But nothing was available.

Does this conclude that the face detection accuracy in project oxford api is way less than face++ and opencv, or did I miss anything in this comparison? 

[1] https://social.msdn.microsoft.com/Forums/getfile/867475

[2] https://social.msdn.microsoft.com/Forums/getfile/867476

[3] https://www.microsoft.com/cognitive-services/en-us/face-api

[4] http://www.faceplusplus.com/demo-detect/

P.S.: Sorry for posting the links separately. My account is not yet verified. 

解决方案

Hi,

 The most possible reason is that the face scale is smaller than 36x36, the smallest detectable face scale in Face-Detection API, please refer https://dev.projectoxford.ai/docs/services/563879b61984550e40cbbe8d/operations/563879b61984550f30395236 for more details.

In order to detect the faces out, you could scale up the image before calling this API.

Thanks for using our service.

这篇关于将Oxford-Face Api与Face ++ Api和Open CV进行比较的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！