Abstract: Recently active learning has attracted a lot of attention in computer vision field, as it is time and cost consuming to prepare a good set of labeled images for vision data analysis. Most existing active learning approaches employed in computer vision adopt most uncertainty measures as instance selection criteria. Although most uncertainty query selection strategies are very effective in many circumstances, they fail to take information in the large amount of unlabeled instances into account and are prone to querying outliers. In this paper, we present a novel adaptive active learning approach that combines an information density measure and a most uncertainty measure together to select critical instances to label for image classifications. Our experiments on two essential tasks of computer vision, object recognition and scene recognition, demonstrate the efficacy of the proposed approach.
Introduction
Active learning (AL) has emerged as a critical technique in computer vision to mitigate the high cost and time required for manual data labeling. Traditional AL methods primarily rely on uncertainty measures to select instances for labeling. However, these "myopic" strategies often fail by ignoring the underlying distribution of unlabeled data, making them susceptible to the "outlier trap"—querying instances that are difficult to classify but offer little representational value for the broader dataset. This paper introduces a robust, adaptive strategy to balance model confusion with data representativeness.
Methodology
The proposed framework integrates two distinct measures into a single query selection criterion:
- Uncertainty Measure ($f(x)$): This uses conditional entropy to identify images where the current classifier (Logistic Regression) is most unsure.
- Information Density Measure ($d(x)$): This measures the "mutual information" a candidate image shares with the rest of the unlabeled pool using a Gaussian Process (GP) framework. An image has high density if it is highly representative of the data distribution.
- Adaptive Combination: The core novelty is the dynamic combination: $h_\beta(x) = f(x)^\beta \cdot d(x)^{1-\beta}$. Instead of a fixed weight, the algorithm selects the optimal $\beta$ in each iteration by simulating the expected classification error across the entire unlabeled set. To maintain computational feasibility, a fast matrix inversion algorithm is used to update the GP.
Experimental Setup
The authors evaluated their method on three standard benchmarks:
- 13 Natural Scene Categories: 3,859 images of natural and man-made scenes.
- Caltech 101: A subset of 3,060 images for object recognition.
- Pascal VOC 2007: 2,989 images for object recognition in "wild" settings.
Results and Discussion
Across all datasets, the proposed approach consistently outperformed standard baselines, including Random Sampling and pure Uncertainty Sampling. By incorporating information density, the algorithm ensures that queried images are both difficult for the model and important to the dataset, effectively avoiding outliers.
A key finding is the advantage of dynamic balancing. As shown in the sensitivity analysis (Figure f), the "optimal" balance between uncertainty and density is not static; the adaptive $\beta$ selection allows the algorithm to shift focus automatically as training progresses, outperforming fixed-combination methods.
Conclusion
The proposed adaptive active learning framework provides a statistically sound method for reducing labeling effort in image classification. While Gaussian Processes introduce computational scaling challenges for massive datasets, the gains in accuracy and label efficiency are significant. Future work may involve exploring approximation methods for larger scales and extending the framework to tasks like object detection.
Summary updated