OpenCV 3.4中HOG的detectMultiScale()/detect()函数异常问题求助

Hey David, let's break down your problems step by step—you're hitting a few common pitfalls when converting SVM models to OpenCV's HOG detector, plus some confusion around confidence scores. Here's what's going on and how to fix it:

1. Why Your Detection Results Are Broken (Center Box vs. Full-Frame Boxes) #

The core issue is a mistake in how you're converting your trained SVM into a HOG detector. OpenCV's HOG detector expects a specific format for the SVM decision function, and your code is missing critical components:

The SVM Decision Function 101 #

For a linear SVM (the standard for HOG-based detection), the decision function looks like this:

f(x) = sum(alpha_i * y_i * x_i · x) - rho

Where:

• alpha_i: Lagrange multipliers for each support vector
• y_i: Label of the support vector (+1 for positive samples, -1 for negative)
• x_i: Support vector features
• rho: The SVM's bias term

Your Conversion Code Is Missing Key Steps #

In your current code, you're just copying the support vectors directly into the HOG detector, without multiplying by alpha_i * y_i (OpenCV actually encodes the y_i sign into the alpha values, so you just need to multiply alpha with the support vectors). Here's the corrected conversion code:

// Get support vectors and SVM decision function parameters
cv::Mat sv = svm->getSupportVectors();
const int sv_total = sv.rows;
cv::Mat alpha, svidx;
double rho = svm->getDecisionFunction(0, alpha, svidx);

// Critical: Multiply support vectors by alpha values (alpha already includes y_i sign)
cv::Mat alpha_mat = alpha.t(); // Convert alpha from column to row vector
cv::Mat detector_mat = alpha_mat * sv; // Compute sum(alpha_i * x_i)

// Build the HOG detector vector
std::vector<float> hog_detector(detector_mat.cols + 1);
memcpy(&hog_detector[0], detector_mat.ptr<float>(), detector_mat.cols * sizeof(float));
hog_detector[detector_mat.cols] = static_cast<float>(-rho);

// Ensure winSize matches exactly what you used during training!
hog.winSize = cv::Size(64, 48);
hog.setSVMDetector(hog_detector);

Also, double-check that your training HOG winSize matches the detection winSize—if they don't, the feature dimensions won't align, and the detector will produce garbage results.

Why detect() vs. detectMultiScale Behave Differently #

• detect() only checks fixed-size windows (your winSize). With a broken detector, every window is triggering the positive threshold (0.0), hence full-frame boxes.
• detectMultiScale() uses image pyramids to check different sizes. The center box you're seeing is just a coincidence where the scaled window's features happened to match the broken detector's incorrect threshold.

2. What Those 20-30 Confidence Values Mean & How to Normalize Them #

The confidence scores you're getting are the raw output of the SVM's decision function f(x):

• Values > 0 mean the model thinks the window is a positive target; higher values = higher confidence.
• Your consistent 20-30 range confirms the detector is broken—normal behavior would have positive targets with high scores, and background windows with scores < 0 or close to 0.

Normalizing Confidence Scores #

Once you fix the detector, if you want to convert these raw scores to a [0,1] range (for easier interpretation), use a sigmoid function:

double normalize_confidence(double raw_score) {
    return 1.0 / (1.0 + exp(-raw_score));
}

This will map large positive scores to ~1, negative scores to ~0, and mid-range scores to values in between.

3. Quick Additional Fixes to Improve Detection #

• Validate your SVM first: Test it on your small, fixed-size test set using svm->predict(test_feature, true)—the second parameter gives you the raw decision score. Confirm positive samples have positive scores, negatives have negative scores.
• Tweak detectMultiScale parameters:
  • scale: Use 1.1 instead of 1.05 (smaller scales add unnecessary computation)
  • minNeighbors: Set to 3-5 to filter duplicate/noise boxes
  • hitThreshold: Instead of 0.0, use a threshold based on your validation set (e.g., the minimum decision score of true positive samples) to filter low-confidence detections.

内容的提问来源于stack exchange，提问作者David W. Romero