Abstract
We propose new techniques for 2-D shape/contour completion, which is one of the important research topics related to shape analysis and computer vision, e.g. the detection of incomplete objects due to occlusion and noises. The purpose of shape completion is to find the optimal curve segments that fill the missing contour parts, so as to acquire the best estimation of the original complete object shapes. Unlike the previous work using local smoothness or minimum curvature priors, we solve the problem under a Bayesian formulation taking advantage of global shape prior knowledge. With the priors, our methods are expert in recovering significant shape structures and dealing with large occlusion cases. There are two different priors adopted in this paper: (i) A generic prior model that prefers minimal global shape transformation (including non-rigid deformation and affine transformation with respect to a reference object shape) of the recovered complete shape; and (ii) a class-specific shape prior model learned from training examples of an object category, which prefers the reconstructed shape to follow the learned shape variation models of the category. Efficient contour completion algorithms are suggested corresponding to the two types of priors. Our experimental results demonstrate the advantage of the proposed shape completion approaches compared to the existing techniques, especially for objects with complex structure under severe occlusion.
Introduction
In computer vision, objects are rarely seen in their entirety. Occlusion, noise, and poor lighting frequently result in fragmented or incomplete contours. To recognize or segment these objects, a system must "fill in the gaps." Historically, researchers relied on local priors, such as the "elastica" model or Euler spirals. These methods assume that the missing curve should simply be the smoothest possible path between two points. While this works for simple geometric shapes, it fails for complex objects like a horse’s legs or a giraffe's neck because local methods lack global context—they do not understand that the curve they are completing is part of a specific, meaningful entity.
Methodology: Bayesian Shape Completion
We propose a Bayesian approach to infer the most likely complete shape $y$ given an incomplete observation $x$. We seek to maximize the posterior $P(y|x) \propto P(x|y)P(y)$, where $P(x|y)$ is the likelihood and $P(y)$ is the shape prior. We introduce two distinct types of global priors:
- Generic Shape Database (GSD): This approach uses a single reference shape for an object class. The algorithm performs partial shape matching (using TPS-RPM) to find where the observed fragment fits on the reference. It then estimates a global transformation—including both affine and non-rigid warping—to reconstruct the missing parts.
- Class-specific Shape Database (CSD): This statistical approach uses multiple training examples of a category to learn an Active Shape Model (ASM). Using Principal Component Analysis (PCA), the system identifies primary shape variations. Completion then becomes a matter of finding parameters in the PCA space that best fit the observed fragments while staying within the "legal" variations of that object type.
Experimental Results
We evaluated our methods using the ETHZ Shapes and Weizmann Horses benchmarks. To simulate real-world challenges, we created synthesized occlusion datasets by randomly placing ellipse-shaped masks over test images.
The quantitative improvements were dramatic. We measured the root-mean-square error (RMSE) of the completed shapes against ground truth. For complex shapes like "Giraffes," the traditional Euler spiral method had an error of 0.9689, while the CSD-based method reduced this to 0.1152. Even for simpler shapes like "Bottles," the error dropped from 0.3392 to 0.0281.
Conclusion
Our research demonstrates that global knowledge of an object's structure is far more powerful than local geometric constraints. By using global priors, we can recover complex structures even under severe occlusion. While current models struggle with large articulations (e.g., a horse jumping), the Bayesian framework provides a robust path forward for shape analysis in cluttered environments. Future work will focus on integrating articulation models to improve robustness to dynamic poses.
Summary updated