ARTOS - Adaptive Real-Time Object Detection System

Björn Barz, Erik Rodner, Joachim DenzlerarXiv 2014 · arXiv

Introduction

The Adaptive Real-Time Object Detection System (ARTOS) is an open-source framework designed to bridge the gap between complex computer vision research and practical, real-world applications. While object detection has seen significant algorithmic advances, deploying customized detectors remains a labor-intensive process often reserved for experts. ARTOS addresses this by providing an end-to-end system that automates training from large-scale datasets like ImageNet and enables rapid "in-situ" adaptation to specific environments through a user-friendly graphical interface.

The Challenge of Custom Object Detection

Traditionally, building a custom object detector involves three major hurdles:

  1. Manual Data Collection: Annotating thousands of images is time-consuming.
  2. Computational Cost: Training complex models can take days.
  3. Domain Shift: Models trained on generic web images often fail in specific deployment settings (e.g., a specific lab or warehouse) due to variations in lighting, background, and camera perspective.

ARTOS overcomes these challenges by automating data acquisition, utilizing a "fast learning" mathematical approach, and offering a streamlined interface for on-the-fly model refinement.

System Methodology

ARTOS synthesizes several state-of-the-art techniques into a cohesive pipeline optimized for speed and usability:

Fast Learning via Linear Discriminant Analysis (LDA)

A core innovation of ARTOS is the use of LDA instead of traditional Support Vector Machines (SVMs). By precomputing background statistics and covariance matrices offline, the system transforms the "training" phase into a simple linear algebra operation. When a user introduces a new object, the system only needs to process a few positive samples, allowing for near-instantaneous model generation.

Feature Representation and Mixture Models

The system employs Whitened Histograms of Orientations (WHO)—decorrelated HOG features that offer increased robustness. To handle varying aspect ratios and viewpoints, ARTOS uses a two-stage clustering process (aspect ratio followed by visual features) to create a "mixture" of detector components. Optimal detection thresholds for these mixtures are determined using a "Harmony Search" heuristic to maximize the F1-score.

The ARTOS Model Catalogue GUI allows users to manage, rename, and adapt various object detectors.

Real-Time Inference

For detection, ARTOS integrates the Fast Fourier Linear Detector (FFLD) library. By leveraging the Convolution Theorem, FFLD significantly accelerates template matching, enabling the system to operate at near real-time speeds suitable for robotics.

In-Situ Adaptation and User Experience

The "in-situ" adaptation feature is a cornerstone of the ARTOS workflow. Users can capture images of an object in its actual deployment environment and immediately incorporate them into the model. This process combines generic knowledge from ImageNet with local context, which empirical results show is the most effective strategy for high-precision detection. The entire system is wrapped in a Python-based GUI (PyARTOS), making high-performance vision tools accessible to non-experts, such as roboticists.

Experimental Evaluation

The system was evaluated using the Office dataset, a standard benchmark for domain adaptation containing common objects in diverse environments. ARTOS was compared against a baseline system, "raptor" (Göhring et al., 2014).

Key Results

As shown in the table below, ARTOS consistently outperformed the baseline. A model trained solely on ImageNet achieved a mean Average Precision (mAP) of 55.5%. However, the most significant gain occurred when this generic model was combined with local "in-situ" data, pushing the precision to 63.5%.

Comparison of mean Average Precision (mAP) between the baseline (raptor) and ARTOS across different training configurations.

Limitations and Future Directions

While ARTOS significantly lowers the barrier to entry for object detection, some limitations remain. The current implementation requires a local copy of the ImageNet dataset, though future versions aim to fetch only necessary data on demand. Additionally, while HOG/WHO features are efficient, modern Deep Learning (CNN) approaches offer higher raw accuracy, albeit at a much higher training cost. Future work includes the development of a public "model catalogue" for community sharing of pre-trained detectors.

Summary updated

Abstract

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.
BibTeX
@misc{barz2014artos,
  author = {Björn Barz and Erik Rodner and Joachim Denzler},
  title = {{ARTOS - Adaptive Real-Time Object Detection System}},
  year = {2014},
  eprint = {1407.2721},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1407.2721},
}
Figure