Azure Custom Vision V3.0训练API精度/召回偏差值始终为0的疑问

Is this a user error? #

Absolutely not! The precisionStdDeviation and recallStdDeviation values showing as 0.0 is not a result of incorrect operation on your part.

Here's why: The Custom Vision V3.0 Training API's GetIterationPerformance endpoint currently doesn't calculate these standard deviation metrics by default. Standard deviation for precision/recall requires running multiple training iterations or cross-validation splits to generate variance data, and the current implementation of this API doesn't include that computation logic. The 0.0 value is the default return when no variance data is available, which is expected behavior right now.

Looking at your sample output, all the standard deviation fields are 0.0—this aligns perfectly with the API's current design:

{
  "averagePrecision": 0.92868346,
  "perTagPerformance": [
    {
      "averagePrecision": 0.4887446,
      "id": "uuid1",
      "name": "tag_name_1",
      "precision": 0.0,
      "precisionStdDeviation": 0.0,
      "recall": 0.0,
      "recallStdDeviation": 0.0
    },
    {
      "averagePrecision": 1.0,
      "id": "uuid2",
      "name": "tag_name_2",
      "precision": 0.0,
      "precisionStdDeviation": 0.0,
      "recall": 0.0,
      "recallStdDeviation": 0.0
    },
    {
      "averagePrecision": 0.9828302,
      "id": "uuid3",
      "name": "tag_name_3",
      "precision": 1.0,
      "precisionStdDeviation": 0.0,
      "recall": 0.5555556,
      "recallStdDeviation": 0.0
    }
  ],
  "precision": 0.9859485,
  "precisionStdDeviation": 0.0,
  "recall": 0.3752228,
  "recallStdDeviation": 0.0
}

Are there plans to enable these statistics? #

As of now, there's no public announcement from the Azure Custom Vision team about adding support for precision/recall standard deviation calculations in the GetIterationPerformance endpoint. If this feature is critical for your use case, you can submit a feature request through Azure's official feedback channels—product teams regularly review user feedback to prioritize new features.

内容的提问来源于stack exchange，提问作者jtlz2