Abstract
Standard object detection descriptors often struggle with complex variations in pose and illumination. While co-occurrence features offer superior descriptiveness, they are prone to overfitting and high computational demands. This paper presents localized co-occurrence features (CoHaar, CoLBP, and CoHOG) and the Generalization and Efficiency Balanced (GEB) framework to address these challenges. By incorporating generalization and efficiency penalties into the RealAdaBoost training process, the GEB framework selects robust and fast features. Experimental results demonstrate that the proposed method achieves a 15% average miss rate on the INRIA dataset and improves detection speed by nearly 40% compared to standard co-occurrence approaches.
Introduction
Standard features such as Haar-like wavelets, Local Binary Patterns (LBP), and Histograms of Oriented Gradients (HOG) frequently fail to detect "hard" samples—objects characterized by significant variations in pose, illumination, or viewpoint. While co-occurrence features, which analyze relationships between pairs of local patterns, provide much higher descriptive power, they introduce two primary challenges: extreme computational complexity and a tendency toward overfitting. The high dimensionality of these features often leads to the "memorization" of training noise rather than the learning of generalizable object characteristics. This research aims to bridge this gap by developing a detector that is both powerful enough for complex objects and efficient enough for real-world applications.
Methodology
Localized Co-occurrence Features
The authors define co-occurrence based on pixel pairs $(a, b)$ separated by a specific spatial offset $(U, V)$. This localized approach is applied to three classic descriptors:
- CoHaar: Captures relationships between Haar-like responses in both intensity and gradient domains.
- CoLBP: Utilizes a simplified version of LBP (uniform patterns) to describe texture relationships.
- CoHOG: Captures the co-occurrence of gradient orientations between localized pixel pairs.
The Generalization and Efficiency Balanced (GEB) Framework
The core innovation of this work is the GEB framework, which modifies the standard RealAdaBoost training process. Typically, boosting algorithms select features solely to minimize classification error ($Z$). The GEB framework introduces two critical penalty terms to the selection criteria:
- Generalization Penalty: This term penalizes features where pixel pairs are spatially distant. The rationale is that proximal pixels maintain more stable contextual relationships, whereas distant pairs are more susceptible to background noise.
- Efficiency Penalty: This term prioritizes computationally "cheaper" features (like CoHaar) during the early stages of the detector cascade, reserving "expensive" features (like CoHOG) for later stages where the classification task is more difficult.
Experimental Results
The proposed method was evaluated on the INRIA Pedestrian, Caltech Pedestrian, and PASCAL VOC 2007 datasets.
Accuracy and Generalization
On the Caltech dataset, the GEB framework improved the accuracy of single co-occurrence features by at least 3%. The combined "All CoFeatures+GEB" detector achieved a 15% average miss rate on the INRIA dataset, outperforming many previous boosting-based methods. Furthermore, on the PASCAL VOC 2007 benchmark, the features achieved a mean Average Precision (mAP) of 43.7%, demonstrating robust generalization across 20 different object categories.
Efficiency Gains
By strategically balancing feature selection, the GEB framework significantly reduced detection time. For a 640x480 image, average detection time dropped from 82.2ms to 50.8ms—a nearly 40% improvement in efficiency without a loss in accuracy.
Conclusion
This research demonstrates that spatial proximity is a key factor in feature robustness; by mathematically favoring local pairs, the model achieves superior generalization. The GEB framework effectively treats computation as a budget, using efficient features to filter easy background windows and saving complex features for difficult samples. While currently limited to single-scale extraction, future work may integrate these co-occurrence relationships into multi-scale feature pyramids or deep learning architectures to further enhance performance.
Summary updated