haar-training

Go application to automate OpenCV Haar Cascade classifier training.

The program orchestrates the full pipeline: downloading negative images from ImageNet, generating positive samples, and training the classifier to produce an XML file usable for object detection.

Prerequisites

• Go (GOPATH mode, no modules)
• OpenCV CLI tools:
  • opencv_createsamples
  • opencv_traincascade

These tools must be available in your PATH.

Installation

go get github.com/marianina8/haar-training
cd $GOPATH/src/github.com/marianina8/haar-training

Go Dependency

• disintegration/imaging — image resizing and grayscale conversion

go get github.com/disintegration/imaging

Usage

go run main.go

The program executes the following steps in order:

1. Download Negative Images

Fetches images from ImageNet (categories: plants, landscapes, crowds, buildings, etc.), resizes them to 200x200 pixels, converts them to grayscale, and stores them in negatives/.

2. Generate bg.txt

Lists all negative image paths in bg.txt, a file required by OpenCV for training.

3. Create Positive Samples

Calls opencv_createsamples to generate variations of the positive image (adidas.png) with rotations and distortions, producing the positives.vec file.

4. Train the Classifier

Calls opencv_traincascade to train the Haar Cascade model from positive and negative samples. The resulting XML classifier is written to the data/ directory.

Project Structure

├── main.go              # Entry point, pipeline orchestration
├── images/
│   └── images.go        # Image downloading and processing from ImageNet
├── samples/
│   ├── create_samples.go      # Positive sample generation (opencv_createsamples)
│   └── create_posvector.go    # Positive vector file creation
├── training/
│   └── cascade.go       # Haar Cascade training (opencv_traincascade)
├── adidas.png           # Example positive image
├── bg.txt               # Negative image list (generated)
├── negatives/           # Downloaded negative images (generated)
├── positives.vec        # Positive sample vector (generated)
└── data/                # Trained classifier output (generated)

Training Parameters

Default parameters in main.go:

Parameter	Value
Sample size	70x70 px
Positive samples	1950 (creation) / 1800 (training)
Negative samples	900
Number of stages	20
Negative image limit	4000
Negative image size	200x200 px

To train on a different object, replace adidas.png with your positive image and adjust the parameters in main.go.