ALCN: Meta-Learning for Contrast Normalization Applied to Robust 3D Pose Estimation

Mahdi Rad, Peter M. Roth, Vincent LepetitarXiv 2017 · arXiv

Abstract

Training deep neural networks to remain robust under varying illumination typically requires massive datasets covering a vast range of lighting conditions. This paper introduces Adaptive Local Contrast Normalization (ALCN), a meta-learning framework that utilizes a "Normalizer" network to predict optimal normalization parameters for local image regions. By learning an adaptive preprocessing step, the system achieves state-of-the-art 3D pose estimation performance even when trained on as few as one to ten images. Our results demonstrate that ALCN significantly outperforms fixed normalization techniques and standard deep architectures in challenging, poorly lit, or specular environments.

Introduction

In computer vision, illumination is a notorious "nuisance variable." Changes in light direction, intensity, or the presence of specular highlights can drastically alter pixel values, often causing deep learning models to fail. The conventional solution is a "brute force" approach: collecting thousands of images of the same object under every conceivable lighting condition. However, for specific tasks like 3D pose estimation of unique industrial parts, creating such datasets is often prohibitively expensive or impossible.

While classical normalization techniques like Difference-of-Gaussians (DoG) or Histogram Equalization exist, they rely on fixed parameters that do not adapt to local image textures. This paper proposes ALCN to bridge this gap by making normalization a learnable, adaptive process.

Methods

The authors propose a meta-learning approach where normalization is treated as a learnable preprocessing step integrated into the detection pipeline. The architecture consists of two primary components:

  1. The Normalizer: A small Convolutional Neural Network (CNN) that analyzes local image windows to predict a set of weights ($w$).
  2. The Model: These weights are applied to a linear combination of ten different 2D Gaussian filters with varying standard deviations. This creates a sophisticated, learned version of the Difference-of-Gaussians method that adapts to the specific content of the image region.
  3. Joint Training: The Normalizer is trained end-to-end alongside a "Detector" CNN. The system learns to select normalization parameters that maximize the Detector's performance, effectively "learning how to see" before "learning what to see."
Overview of the ALCN framework showing the minimal training images required and the resulting robustness in complex test scenarios.

To ensure practical utility, the final normalization is computed efficiently using Hadamard products and convolutions, requiring approximately 50ms on a standard GPU for a 128x128 image.

Datasets and Experimental Setup

The versatility of ALCN was validated across several challenging datasets:

  • Phos Dataset: Used for the initial joint training of the Normalizer, containing various objects under 15 different lighting conditions.
  • ALCN-2D: A new dataset featuring objects with diverse material properties (plastic, velvet, and specular metal) under 1,200 test illuminations and cluttered backgrounds.
  • Box and Duck Sequences: Challenging 3D pose estimation tasks involving drastic perspective changes and unseen lighting conditions.

Key Results and Insights

ALCN demonstrates remarkable data efficiency and robustness. On the ALCN-2D dataset, the method achieved an Area Under the Curve (AUC) of 0.787, significantly higher than the best un-normalized deep network (0.606) and standard DoG normalization (0.611).

Robust 3D pose estimation results across different objects (bee, box, and duck) under varying illumination and cluttered backgrounds.

Data Efficiency

A key finding is that the system provides high-quality results with only one real training image (augmented with synthetic noise) and reaches peak performance with just ten images. This is a significant improvement over standard deep learning models that require thousands of samples to achieve similar invariance.

3D Pose Robustness

In 3D pose estimation tests, a standard VGG network's accuracy dropped to 0% under challenging light changes. By integrating ALCN, the same network maintained 50-76% accuracy, proving that explicit local normalization is more effective than increasing network depth (e.g., ResNet-110) for handling illumination nuisances.

Conclusion

ALCN provides a powerful alternative to brute-force data collection for robust computer vision. By making normalization local and adaptive, the model successfully handles cast shadows, self-shadows, and specular highlights that typically blind traditional filters. Future work will focus on optimizing the Normalizer for high-resolution real-time applications and exploring joint normalization across color channels.

Summary updated

Abstract

To be robust to illumination changes when detecting objects in images, the current trend is to train a Deep Network with training images captured under many different lighting conditions. Unfortunately, creating such a training set is very cumbersome, or sometimes even impossible, for some applications such as 3D pose estimation of specific objects, which is the application we focus on in this paper. We therefore propose a novel illumination normalization method that lets us learn to detect objects and estimate their 3D pose under challenging illumination conditions from very few training samples. Our key insight is that normalization parameters should adapt to the input image. In particular, we realized this via a Convolutional Neural Network trained to predict the parameters of a generalization of the Difference-of-Gaussians method. We show that our method significantly outperforms standard normalization methods and demonstrate it on two challenging 3D detection and pose estimation problems.
BibTeX
@misc{rad2017alcn,
  author = {Mahdi Rad and Peter M. Roth and Vincent Lepetit},
  title = {{ALCN: Meta-Learning for Contrast Normalization Applied to Robust 3D Pose Estimation}},
  year = {2017},
  eprint = {1708.09633},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1708.09633},
}
Figure