Detection of First and Second Order Motion

Alexander Grunewald, Heiko NeumannNeurIPS 1997

This paper, authored by Alexander Grunewald and Heiko Neumann, presents a biologically inspired computational model explaining how the human visual system detects different types of motion. By mapping specific mathematical operations to known types of brain cells (simple and complex cells), the authors provide a framework for understanding why we can see a moving object even when its average brightness doesn't change.

The Challenge: First vs. Second-Order Motion

In classical computer vision and neuroscience, motion detection is often modeled using linear filters that look for moving "blobs" of light or dark (luminance). This is known as first-order motion. However, humans can also detect second-order motion, where the moving object is defined by changes in texture, flicker, or contrast, even if its average luminance matches the background (e.g., a "camouflage" bar moving across a static noisy background).

Standard linear models, like the Reichardt detector or the Adelson-Bergen motion energy model, fail to detect second-order motion because the "energy" of the signal often points in the wrong direction or cancels out. The challenge is to create a model that handles both types of motion using biologically plausible mechanisms without requiring a complex, high-level "object recognition" stage first.

A Three-Stage Neural Model

The researchers developed a three-stage computational model to address this:

  1. Unoriented Stage: The input (a space-time matrix of light) is processed by "ON" and "OFF" filters (Difference of Gaussians). This mimics the retina and Lateral Geniculate Nucleus (LGN), creating separate channels for light increments and decrements.
  2. Phase Insensitive Stage (The "Complex Cell" Pathway): This pathway adds the ON and OFF signals together before filtering them for motion. This pooling acts as a "full-wave rectification," making the detector sensitive to the presence of a feature regardless of whether it is bright or dark. This allows it to track the "envelope" of second-order motion.
  3. Phase Sensitive Stage (The "Simple Cell" Pathway): This pathway keeps ON and OFF signals separate. It is highly tuned to specific light/dark arrangements. While excellent at detecting first-order motion, it fails on second-order stimuli because the reversing contrast polarities "confuse" the filter.

The model also includes a competitive inhibition mechanism where the phase-insensitive pathway inhibits the phase-sensitive one to prevent false signals.

Key Results and Insights

Rather than relying on large-scale external datasets, the authors used synthetic stimuli designed to test specific motion types: a bright bar moving across a dark background (first-order) and a "contrast-reversing" bar where pixels flip between bright and dark as the bar moves (second-order).

Model performance on first and second-order motion stimuli
  • Detection Success: As shown in the results, the model successfully detected rightward motion for both first-order (luminance) and second-order (contrast-reversing) bars in the phase-insensitive pathway.
  • Pathway Specificity: The phase-sensitive pathway (simple cells) correctly identified first-order motion but remained silent or produced poor signals for second-order motion, matching human psychophysical performance.
  • Parallel Processing: Rather than second-order motion being a "higher-level" process that happens after first-order detection, the model suggests they happen in parallel.
  • Challenging the "Hierarchical" View: Traditionally, it was thought that complex cells are built by combining outputs from simple cells. This paper suggests that for second-order motion detection to work, some complex cells must receive direct input from the LGN (unoriented stage), bypassing simple cells.

Limitations and Future Work

The current model is 1D in space (plus 1D in time). A 2D spatial extension would be necessary to fully account for orientation selectivity in a more realistic visual environment. Furthermore, the inhibition between pathways is simplified; real cortical circuits likely involve more complex feedback loops. Finally, while the model predicts direct LGN-to-complex-cell connections, the authors acknowledge that many complex cells do receive simple cell input, highlighting that this model specifically addresses the subset of cells required to explain second-order motion.

Summary updated

BibTeX
@inproceedings{grunewald1997detection,
  author = {Alexander Grunewald and Heiko Neumann},
  title = {{Detection of First and Second Order Motion}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1997},
}
Figure