The easiest way to perform cross-validation in OpenCV is to do the data splits by hand.

For example, to implement two-fold cross-validation, we would perform the following procedure:

1. Load the dataset:

      In [1]: from sklearn.datasets import load_iris
      ...     import numpy as np
      ...     iris = load_iris()
      ...     X = iris.data.astype(np.float32)
      ...     y = iris.target

2. Split the data into two equally sized parts:

      In [2]: from sklearn.model_selection import model_selection
      ...     X_fold1, X_fold2, y_fold1, y_fold2 = train_test_split(
      ...         X, y, random_state=37, train_size=0.5
      ...     )

3. Instantiate the classifier:

      In [3]: import cv2
      ...     knn = cv2.ml.KNearest_create()
      ...     knn.setDefaultK(1)

4. Train the classifier on the first fold, then predict the labels of the second fold:

      In [4]: knn.train(X_fold1, cv2.ml.ROW_SAMPLE, y_fold1)
      ...     _, y_hat_fold2 = knn.predict(X_fold2)

5. Train the classifier on the second fold, then predict the labels of the first fold:

      In [5]: knn.train(X_fold2, cv2.ml.ROW_SAMPLE, y_fold2)
      ...     _, y_hat_fold1 = knn.predict(X_fold1)

6. Compute accuracy scores for both folds:

      In [6]: from sklearn.metrics import accuracy_score
      ...     accuracy_score(y_fold1, y_hat_fold1)
      Out[6]: 0.92000000000000004
      In [7]: accuracy_score(y_fold2, y_hat_fold2)
      Out[7]: 0.88

This procedure will yield two accuracy scores, one for the first fold (92% accuracy) and one for the second fold (88% accuracy). On average, our classifier hence achieved 90% accuracy on unseen data.