Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Study of Shape Modeling Against Noise

Cheng Long, Adrian Barbu

TL;DR

The paper studies shape denoising for binary shape representations, introducing six realistic noise types to perturb shapes and using Intersection over Union (IoU) as an objective recovery metric. Shapes are represented as binary images of size $128 \times 128$, aligned to a centered, normalized form, enabling evaluation across diverse denoising models. Seven methods, including ASM, DBM, CDBM, EBM, U‑Net, Deeplabv3+, and MAE, are evaluated on the Weizmann Horse and Flycatcher datasets, with MAE and U‑Net delivering the strongest denoising performance across noise types and EBM excelling for real-image noise. The work provides a modality-agnostic, IoU-based benchmark for shape denoising and lays groundwork for applying strong shape denoisers to color-image segmentation tasks.

Abstract

Shape modeling is a challenging task with many potential applications in computer vision and medical imaging. There are many shape modeling methods in the literature, each with its advantages and applications. However, many shape modeling methods have difficulties handling shapes that have missing pieces or outliers. In this regard, this paper introduces shape denoising, a fundamental problem in shape modeling that lies at the core of many computer vision and medical imaging applications and has not received enough attention in the literature. The paper introduces six types of noise that can be used to perturb shapes as well as an objective measure for the noise level and for comparing methods on their shape denoising capabilities. Finally, the paper evaluates seven methods capable of accomplishing this task, of which six are based on deep learning, including some generative models.

A Study of Shape Modeling Against Noise

TL;DR

The paper studies shape denoising for binary shape representations, introducing six realistic noise types to perturb shapes and using Intersection over Union (IoU) as an objective recovery metric. Shapes are represented as binary images of size , aligned to a centered, normalized form, enabling evaluation across diverse denoising models. Seven methods, including ASM, DBM, CDBM, EBM, U‑Net, Deeplabv3+, and MAE, are evaluated on the Weizmann Horse and Flycatcher datasets, with MAE and U‑Net delivering the strongest denoising performance across noise types and EBM excelling for real-image noise. The work provides a modality-agnostic, IoU-based benchmark for shape denoising and lays groundwork for applying strong shape denoisers to color-image segmentation tasks.

Abstract

Shape modeling is a challenging task with many potential applications in computer vision and medical imaging. There are many shape modeling methods in the literature, each with its advantages and applications. However, many shape modeling methods have difficulties handling shapes that have missing pieces or outliers. In this regard, this paper introduces shape denoising, a fundamental problem in shape modeling that lies at the core of many computer vision and medical imaging applications and has not received enough attention in the literature. The paper introduces six types of noise that can be used to perturb shapes as well as an objective measure for the noise level and for comparing methods on their shape denoising capabilities. Finally, the paper evaluates seven methods capable of accomplishing this task, of which six are based on deep learning, including some generative models.
Paper Structure (8 sections, 7 equations, 2 figures, 2 tables)

This paper contains 8 sections, 7 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. The Shape Denoising Problem
  4. Shape Representation
  5. Introducing Noise in Shapes
  6. Shape Denoising Evaluation
  7. Evaluation of Shape Denoising Methods
  8. Conclusion

Figures (2)

  • Figure 1: Shape denoising example. The noisy shape (b) has been obtained from the original shape (a) by a noise inducing process such as those described in Section \ref{['sec:noise']}. A shape denoising method is used to obtain the denoised shape (c).
  • Figure 2: Illustration of shape alignment and the six types of shape noise introduced in this work. (a) Shape before alignment, (b) Shape after alignment, (c) Salt and pepper noise, (d) Circle noise, (e) Real image noise, (f) Occlusion noise, (g) Thresholded probability noise, (h) Detection image noise.