Abstract: Traditional place categorization approaches in robot vision assume that training and test images have similar visual appearance. Therefore, any seasonal, illumination and environmental changes typically lead to severe degradation in performance. To cope with this problem, recent works have proposed to adopt domain adaptation techniques. While effective, these methods assume that some prior information about the scenario where the robot will operate is available at training time. Unfortunately, in many cases this assumption does not hold, as we often do not know where a robot will be deployed. To overcome this issue, in this paper we present an approach which aims at learning classification models able to generalize to unseen scenarios. Specifically, we propose a novel deep learning framework for domain generalization. Our method develops from the intuition that, given a set of different classification models associated to known domains (e.g. corresponding to multiple environments, robots), the best model for a new sample in the novel domain can be computed directly at test time by optimally combining the known models. To implement our idea, we exploit recent advances in deep domain adaptation and design a Convolutional Neural Network architecture with novel layers performing a weighted version of Batch Normalization. Our experiments, conducted on three common datasets for robot place categorization, confirm the validity of our contribution.
Introduction
Robots must accurately categorize their surroundings—identifying rooms as offices, corridors, or kitchens—to navigate and interact effectively. However, deep learning models for robot vision frequently suffer from domain shift. A model trained on images from a sunny office in one city often fails when deployed in a rainy office in another, due to variations in lighting, furniture, and environmental conditions.
While Domain Adaptation (DA) has been used to mitigate this, it typically requires access to data from the target environment during training. In real-world robotics, we often deploy systems into entirely "unseen" environments. This research addresses Domain Generalization (DG), aiming to build models that work out-of-the-box in brand-new scenarios without prior exposure to target-specific data.
The Domain Generalization Framework
The core intuition of this approach is that a novel environment can be represented as a combination of known environments. Instead of training a single rigid model, the framework leverages multiple classification models associated with known domains (e.g., different buildings or robots). At test time, the system dynamically computes the best model for a new sample by optimally combining these known models.
Deep Architecture with Weighted Batch Normalization
To implement this "blending" of knowledge, the authors introduce two primary architectural innovations within a Convolutional Neural Network (CNN):
- Weighted Batch Normalization (WBN): Standard Batch Normalization (BN) uses global statistics (mean and variance) for an entire dataset. WBN maintains separate statistics for different source domains. At test time, it creates a weighted combination of these statistics to normalize the features of the current image, effectively synthesizing a "virtual domain" that matches the unseen environment.
- The Lateral Branch: To determine the appropriate weights for WBN, a small lateral network branch is added. This branch estimates the probability that an incoming image from an unknown environment belongs to each of the known training domains. These probabilities serve as the weights ($w_1, ..., w_N$) for the WBN layers.
Additionally, the framework supports Latent Domain Discovery. If explicit domain labels (like "Building A" or "Nighttime") are unavailable, the network can automatically cluster training data into latent domains, discovering its own environmental categories to improve normalization.
Experimental Evaluation
The approach was validated on three major robot vision benchmarks:
- COLD (COsy Localization Database): Testing generalization across three labs under varying lighting (Cloudy, Sunny, Night). The ResNet-based WBN model achieved 40.0% average accuracy, outperforming the standard baseline of 37.5%.
- VPC (Visual Place Categorization): Focusing on room types across six different houses. The method reached 56.5% accuracy, a significant improvement over previous non-deep learning state-of-the-art methods like CENTRIST and HOUP (~50%).
- SPED (Specific PlacEs Dataset): A large-scale outdoor dataset testing seasonal generalization. The WBN model showed a 4% improvement in the challenging task of generalizing from August data to February data.
Conclusion and Insights
The results demonstrate that soft assignment is key; by blending statistics from multiple domains, the model can adapt to the nuances of unseen locations. Furthermore, this work reinforces that the Batch Normalization layer is a highly effective site for handling domain shifts by aligning new environment features with the classifier's training distribution.
Future work may focus on automatically determining the optimal number of latent domains and expanding source data to include synthetic or web-scraped images, further hardening robotic systems against the unpredictability of real-world deployments.
Summary updated