Abstract
This paper presents a Bayesian framework for object recognition that integrates local descriptors with their spatial relationships. By accounting for descriptor variability, saliency, and spatial coherence, the model provides a robust mechanism for identifying objects in cluttered or occluded scenes. Experimental results on a large database demonstrate significant performance gains over traditional voting-based methods.
Introduction
In the late 1990s, computer vision researchers sought to recognize objects in complex, real-world scenes. Early methods were often divided between global descriptors, which were sensitive to occlusion, and local descriptors, which could be overwhelmed by background noise. The prevailing "voting" schemes of the time lacked a formal mathematical foundation to handle uncertainty or distinguish between common and rare features. This work addresses these limitations by introducing a structured probabilistic model that formally integrates feature similarity, saliency, and spatial layout.
Methodology
The proposed Bayesian model calculates the probability of a model image ($M_m$) being present in a scene given a set of hypotheses ($\mathcal{H}$). The probability is decomposed into three independent components:
- Global Coherence ($P_h$): Based on the principle of parsimony, this component favors hypotheses that explain the scene using the fewest number of objects.
- Quality of Correspondences ($P_e$): Evaluates matches based on Mahalanobis distance (similarity) and saliency. Saliency weights descriptors by their uniqueness across the database, effectively filtering out common, uninformative features.
- Spatial Coherence ($P_s$): Validates the geometric arrangement of local features using local angular configurations, ensuring the spatial layout in the query matches the database model.
To manage computational complexity, the system employs a heuristic to evaluate the top 2,000 most promising hypotheses from a vast search space.
Experimental Setup
The model was validated using a diverse database of 1,020 images, including 200 paintings, 100 aerial photographs, and 720 images of 3D objects from the Columbia database.
Testing involved challenging scenarios with varying viewpoints (up to 20 degrees), lighting changes, and significant clutter.
Results and Discussion
The probabilistic approach demonstrated superior robustness compared to traditional voting methods. A key finding was the model's performance under occlusion and reduced query size.
As shown in Figure 2, the "proba" method maintains a high recognition rate even when the query image is reduced to 20% of its original size, whereas the voting method's accuracy drops sharply. Furthermore, the probabilistic model produces much clearer distinctions between the correct match and false positives; the top match often has a score significantly higher than the second-best match, whereas voting scores tend to be much closer together.
Conclusion
By decoupling spatial layout from local appearance and formally integrating saliency, this framework provides a modular and efficient solution for instance-level recognition. Future work will focus on extending this model to object class recognition, which requires handling higher intra-class variability, and using saliency to prune uninformative data from large-scale databases.
Summary updated