Dlib-ImageNet-Datasets

Preprocessed Stable ImageNet-1K datasets for efficient computer vision prototyping with Dlib.

📌 Overview

This repository provides ready-to-use ImageNet-1K datasets preprocessed in multiple resolutions (32×32 to 256×256) for the Dlib machine learning library. Designed for rapid experimentation, benchmarking, and model training, these datasets eliminate preprocessing overhead while ensuring consistency across experiments.

🚀 Immediately available:

• A ready-to-use 32×32 resolution dataset (ideal for lightweight model prototyping) in the /dataset directory.

🛠️ Flexible generation - The included C++14 tool lets you create custom datasets in any resolution (e.g., 64×64, 128×128, 224×224, etc.) from raw ImageNet-1K sources. Perfect for:

• Rapid experimentation
• Resolution-impact benchmarking
• Consistent model training pipelines

🔧 Usage - The raw "Stable ImageNet-1K" images can be downloaded from:
→ Kaggle Dataset

1. Download and extract the compressed file locally
2. Point the extraction root directory to our processing tool:

./create_dataset path/to/extracted_folder output_dataset.dat 224  # Example for 224x224

Create Custom Datasets

Compile and run the included tool to process raw ImageNet-1K images:

g++ -std=c++14 src/create_dataset.cpp -o create_dataset -ldlib -lpthread
./create_dataset path/to/imagenet_root datasets/128x128/imagenet_128.dat 128

Evaluate Models

Load pre-split training/testing sets:

std::vector<dlib::matrix<rgb_pixel>> train_images, test_images;
std::vector<unsigned long> train_labels, test_labels;
dlib::load_stable_imagenet_1k("datasets/64x64/imagenet_64.dat", train_images, train_labels, test_images, test_labels);