ARTOS -- Adaptive Real-Time Object Detection System

ARTOS is all about creating, tuning, and applying object detection models with just a few clicks. In particular, ARTOS facilitates learning of models for visual object detection by eliminating the burden of having to collect and annotate a large set of positive and negative samples manually and in addition it implements a fast learning technique to reduce the time needed for the learning step. A clean and friendly GUI guides the user through the process of model creation, adaptation of learned models to different domains using in-situ images, and object detection on both offline images and images from a video stream. A library written in C++ provides the main functionality of ARTOS with a C-style procedural interface, so that it can be easily integrated with any other project.

💡 Research Summary

The paper presents ARTOS (Adaptive Real‑Time Object Detection System), an open‑source framework that streamlines the entire workflow of creating, adapting, and deploying object detection models with only a few user interactions. The system builds on established techniques—ImageNet for large‑scale data acquisition, HOG‑based features enhanced by Felzenszwalb’s improvements and the Whitened HOG (WHO) transformation of Hariharan et al., and Linear Discriminant Analysis (LDA) for rapid classifier training. By pre‑computing the negative class mean and covariance matrix, ARTOS reduces the learning step to a simple computation of the positive class mean, allowing near‑instant model generation even when only a few positive samples are available.

Training data are organized through a two‑stage clustering process. First, images are grouped by aspect‑ratio; second, each group is subdivided using k‑means on WHO features. A separate linear detector is trained for each cluster using the LDA closed‑form solution w = Σ⁻¹(μ₊ − μ₋). These detectors are then combined into a mixture model whose final detection score is the maximum of the individual scores. This mixture approach captures intra‑class variability without requiring a complex hierarchical model.

Threshold (bias) selection is addressed in two phases. Individual detector biases are tuned to maximize the F1‑measure on a validation set composed of positive samples and randomly selected negatives from other ImageNet synsets. Because the mixture’s overall score is the maximum of its components, a globally optimal set of biases is essential. The authors employ the Harmony Search heuristic to approximate the optimal bias combination, noting that any reasonable heuristic would likely yield comparable results.

For real‑time inference, ARTOS integrates the Fast Fourier Linear Detector (FFLD) by Dubout and Fleuret, which leverages the convolution theorem to perform template matching in the frequency domain, achieving tens of frames per second on a CPU. This enables immediate deployment on video streams without GPU acceleration.

The user experience is a central focus. A C++ core library is exposed through a C‑style procedural API, wrapped by a Qt‑based graphical interface and a Python front‑end (PyARTOS). Users simply select a synset, adjust optional clustering and threshold parameters, and click “Learn!”. The system automatically downloads (or accesses locally stored) ImageNet images, computes features, clusters, trains detectors, and optimizes thresholds. An “in‑situ” mode lets users capture images from the target environment, annotate them, and add new sub‑models to the existing mixture, thereby compensating for domain shift.

Experimental evaluation follows the protocol of Göh​ring et al. (2014) on the Office dataset. Three configurations are compared: (1) a model trained solely on ImageNet data, (2) a model trained only on in‑situ images, and (3) an adapted mixture that combines both sources. ARTOS achieves F1 scores of 55.5 % (ImageNet only), 51.0 % (in‑situ only), and 63.5 % (adapted mixture). These results surpass the original method’s scores of 49.7 %, 36.6 %, and 54.1 % respectively, demonstrating that the clustering and bias‑optimization steps contribute significantly to performance gains.

Limitations acknowledged by the authors include the current requirement for a local copy of ImageNet, the exclusive reliance on WHO/HOG features (as opposed to modern deep‑learning descriptors), and the lack of comparative studies with alternative optimization strategies. Future work aims to provide an online ImageNet downloader, publish a public model catalog, and explore hybrid pipelines that incorporate convolutional neural network features.

In summary, ARTOS delivers a practical, end‑to‑end solution for large‑scale object detection: it automates data collection, offers a fast LDA‑based learning scheme, supports multi‑component mixture models, provides heuristic bias optimization, and enables real‑time detection through an efficient Fourier‑based engine. Its user‑friendly GUI and Python bindings lower the barrier for both researchers and practitioners to build, adapt, and deploy robust detectors with minimal effort.