Grid Coding: A Preprocessing Technique for Robot and Machine Vision

TL;DR

This paper introduces "grid coding," a novel technique for machine vision that uses controlled illumination to encode scene information directly into a grid pattern. This allows for a more robust and heuristic-free extraction of planar areas from 3D scenes, paving the way for more capable robot vision systems.

What Problem Does This Paper Address?

Traditional machine vision approaches for robots often rely on methods analogous to human visual processing, focusing on detecting edges and then inferring shapes. This process is prone to noise and uncertainty, especially when dealing with complex scenes or subtle photometric differences. A significant challenge is reliably extracting meaningful information, like planar areas, from a 3D scene to build a coherent understanding. The authors highlight that existing methods struggle with scenes where illumination is uniform, making edge detection difficult, and that current approaches often postpone complex analysis to later stages, relying on heuristics. The paper aims to provide a more direct and robust method for decomposing scenes into fundamental geometric components, specifically planar areas.

Methods

The core innovation is grid coding. Instead of relying solely on ambient light and edge detection, the authors propose controlling the illumination itself. This is achieved by projecting a grid pattern onto the scene. The scene is then captured, and the distortion of this grid reveals information about the scene's geometry.

The key idea is to encode the desired information (planar areas) as the modulation on a spatial frequency carrier grid. This transforms the problem of extracting planar areas into a linear frequency domain filtering task. The authors leverage the power of the Fast Fourier Transform (FFT) for efficient processing.

Specifically, they describe two main approaches:

  1. Grid-Coded Illumination: Shining a "grid-coded" light onto the scene. This can be a simple crossed grating, as shown in Figure 2, where different orientations of the grid on different planar surfaces help distinguish them.
  2. Fourier Domain Filtering: After capturing the grid-coded image, its 2D Fourier transform is computed. A planar area, when coded this way, results in a Fourier spectrum with a specific distribution of delta functions. By applying band-pass filters in the frequency domain, these specific components can be isolated, and their inverse transform reconstructs the isolated planar area.

The paper also touches upon "New Filtering Approaches" that use more complex grid structures (like linear zone plates) to encode positional information, enabling more precise 3D position determination through triangulation.

Datasets

The paper does not discuss training or evaluation on standard datasets. Instead, it uses a dodecahedron as a test example (Figure 3). This object was photographed with grid-coded illumination, scanned, and then processed to demonstrate the grid coding technique. The authors mention that experiments were performed both optically and digitally.

Key Insights and Ideas

  • Shifting the Burden to Illumination: The most significant insight is moving the complexity from post-processing (edge detection, fitting lines) to controlled illumination. This makes the extraction process more direct and less susceptible to noise.
  • Fourier Domain as a Natural Fit: Framing scene decomposition as a spatial frequency filtering problem is elegant. The Fourier transform naturally decomposes signals into their constituent frequencies, making it ideal for extracting patterns encoded as modulations.
  • Heuristic-Free Extraction: By encoding information directly, the method aims to be "heuristic-free," meaning it relies on fundamental mathematical operations rather than ad-hoc rules, leading to more reliable results.
  • Beyond Edges: The approach focuses on extracting planar areas first, defining edges as intersections of these planes. This is a more fundamental geometric representation than just edge points.
  • Potential for 3D Information: The paper hints at the ability to extract 3D information, not just 2D scene decomposition, by using more sophisticated grid coding and processing techniques.

Key Results

The paper demonstrates the successful extraction of individual planar components from a grid-coded dodecahedron scene. Figures 6-9 show examples of planes extracted using sector filters in the Fourier domain. The authors state that this decomposition into elementary planar components, including the background, is possible. While quantitative metrics like accuracy percentages are not provided, the qualitative results visually confirm the method's ability to isolate distinct surfaces.

Limitations and Future Work

  • Ambiguity in General Scenes: The authors acknowledge that grid coding, especially with simple crossed grids, does not always provide a unique decomposition of a general scene. Adjacent faces of convex bodies are uniquely coded, but more complex scenarios may require further refinement.
  • Illumination Control: The method relies on controlled illumination, which might not always be feasible in all real-world robot vision applications.
  • Future Work: The paper suggests several promising directions:
  • Stereo Vision: Generating stereoscopic pairs automatically from grid-coded images.
  • Difference Extraction: Applying grid coding to compute differences between images, potentially for video compression or change detection.
  • 3D Recording: Recording full 3D information by moving the camera and grid coding the resulting streaks, where streak length indicates range.
  • More Sophisticated Grids: Using grids that encode positional information for more precise 3D localization.

The authors believe their heuristic interface is pushed deeper into the recognition process, where it "probably belongs," suggesting that the fundamental geometric decomposition is a more appropriate place for robust machine vision.

Summary updated

BibTeX
@inproceedings{1971grid,
  title = {{Grid Coding: A Preprocessing Technique for Robot and Machine Vision}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1971},
}
Figure