Strict Integer as Categorical vs Heuristic Inference in OpenCV TrainData: Which Ensures Consistent Feature Types?

Question

Goal
Determine the most reliable approach for determining variable types in cv::ml::TrainData when a feature column contains a mixture of integer and floating‑point values.
Constraint
OpenCV’s current inference heuristic may classify a mixed column as continuous, potentially misrepresenting categorical data. The library documentation does not specify how such mixed types should be handled, nor does it provide a flag to enforce stricter rules.
Uncertainty
The behavior may vary across OpenCV releases, leading to inconsistent training results. Developers must decide whether to rely on the automatic inference, explicitly set the varType vector, or advocate for a stricter default that treats any integer‑only column as categorical.
Questions

Will OpenCV treat all integer columns as categorical by default if a stricter inference rule is adopted?
Is there a documented parameter or flag that allows users to override the current heuristic without explicitly providing a varType vector?
What impact would enforcing a stricter inference rule have on backward compatibility and existing codebases that rely on the current heuristic?

Zoha · Answer

Answer

The only way to guarantee that a feature column is treated exactly as you intend—categorical or numerical—when OpenCV’s cv::ml::TrainData is built is to supply the varType vector explicitly. The library’s default inference is a heuristic that can mis‑classify mixed or borderline columns, and there is no flag to tighten that rule without code changes.

How the Heuristic Works

All values are integers and the number of distinct integer values is ≤ 20 (default) → treated as CV_VAR_CATEGORICAL.
Any floating‑point value present → the entire column is forced to CV_VAR_NUMERICAL.
Columns that are purely integer but have >20 distinct values are also marked numerical.

These thresholds can shift between OpenCV releases, so relying on the heuristic can produce non‑deterministic training results when the data set changes even slightly.

Why Explicit varType Is Safer

You control the vector; the values you provide are what the model sees.
It works the same across all OpenCV versions (4.5, 4.8, 4.10, etc.).
No hidden thresholds or version‑specific defaults.

Practical Steps

Create your feature matrix (e.g., cv::Mat features;) and target vector.

Define the type vector:

std::vector varType;
// Example: first column categorical, second numerical
varType.push_back(cv::ml::TrainData::VAR_CATEGORICAL);
varType.push_back(cv::ml::TrainData::VAR_NUMERICAL);

Build the TrainData object:

auto trainData = cv::ml::TrainData::create(
    features, cv::ml::ROW_SAMPLE, target,                     // data, layout, labels
    cv::noArray(), cv::noArray(), cv::noArray(),             // missingMask, varIdx, varName
    varType, cv::noArray(), cv::noArray()                     // varType, varIdx, varName
);

Verify the stored types:

for (int i = 0; i < trainData->getVarCount(); ++i) {
    std::cout << "Column " << i << " type: " << trainData->getVarType(i) << "\n";
}

Impact on Existing Code

Adopting a stricter default that treats every integer column as categorical would break backward compatibility. Models that previously relied on the <=20 distinct values rule would suddenly change behavior, potentially altering predictions and training convergence.

Missing Diagnostic Detail

To tailor the vector precisely, it would help to know the exact threshold value used by the OpenCV release you target. Please confirm whether your environment uses the default 20‑value rule or a different one.

Strict Integer as Categorical vs Heuristic Inference in OpenCV TrainData: Which Ensures Consistent Feature Types?

Goal

Constraint

Uncertainty

Questions

1 answer