Incremental Learning Framework for Indoor Scene Recognition

Aram Kawewong, Rapeeporn Pimup, Osamu HasegawaAAAI 2013 · DOI

Introduction

A fundamental challenge in robotics is enabling machines to recognize surroundings dynamically rather than relying on static, pre-trained models. In real-world scenarios, such as a service robot deployed in a new hospital, the system must learn novel indoor locations incrementally through interaction. Indoor scene recognition is particularly complex due to high intra-class variance; some scenes are defined by spatial layout (e.g., corridors), while others are defined by constituent objects (e.g., kitchens). Traditional offline methods suffer from "catastrophic forgetting" or become computationally prohibitive as data accumulates. This paper, presented at AAAI 2013, proposes the $n$-SOINN-SVM framework to balance high accuracy, rapid updates, and category scalability.

Methodology: The $n$-SOINN-SVM Framework

The core of the proposed system is the $n$-value Self-Organizing Incremental Neural Network ($n$-SOINN), which facilitates efficient data compression and incremental learning.

Overview of the incremental learning framework architecture, including the Place Category Model and human-in-the-loop feedback loop.

Data Compression and Representation

Standard SOINN is an unsupervised clustering algorithm that represents data topology through nodes. The authors introduced a parameter ($n$) that dictates the number of "wins" a node must accumulate before spawning a new neighbor. This allows the system to compress thousands of high-dimensional feature vectors (e.g., GIST, S-PACT) into a compact set of representative nodes. This compression preserves the category's "shape" while significantly reducing the computational load for subsequent classification.

Classification and Human Interaction

The framework utilizes a Support Vector Machine (SVM) trained on these representative nodes rather than raw data. By using a standardized Euclidean distance measure, the system accounts for varying feature variances across different scene categories. The framework is designed for "human-in-the-loop" interaction: when the robot encounters uncertainty or makes an error, a human provides a correction. The system then updates its $n$-SOINN and SVM models in roughly two seconds, allowing for near-instantaneous learning.

Experimental Results

The framework was evaluated using the MIT 67-category indoor scene dataset and an "Extended Learning Set" consisting of 1,340 additional images retrieved from Google to simulate real-world feedback.

Accuracy and Efficiency

The $n$-SOINN-SVM framework achieved a state-of-the-art accuracy of 45.0% on the MIT 67 dataset when leveraging extended learning. This performance matches top-tier offline methods that process all data simultaneously.

Comparison of accuracy across various scene recognition methods, highlighting the performance of the proposed n-SOINN-SVM with extended learning.

Computational Speed

A key advantage of the framework is its update speed. While traditional offline baselines required over 100 seconds to re-train on new data, the $n$-SOINN-SVM framework completed updates in approximately 2 seconds. This efficiency is driven by the massive reduction in training samples; for instance, negative samples were reduced from thousands to a few hundred representative nodes with negligible accuracy loss.

Discussion and Key Insights

The success of the framework hinges on the "n" parameter, which controls network density. A low $n$ captures fine-grained details for positive samples, while a higher $n$ creates generalized models for negative samples. While the system excels in interactive environments, its performance in purely autonomous "solo" operations remains a target for future optimization. Integrating modern deep-learning features into this incremental pipeline is expected to further enhance baseline accuracy.

Summary updated

BibTeX
@inproceedings{kawewong2013incremental,
  author = {Aram Kawewong and Rapeeporn Pimup and Osamu Hasegawa},
  title = {{Incremental Learning Framework for Indoor Scene Recognition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2013},
  doi = {10.1609/aaai.v27i1.8584},
}

Loading related content...

Figure