Animate Vision in a Rich Environment

Tomas Uhlin, Jan-Olof EklundhIJCAI 1995

Introduction

In the mid-1990s, researchers at KTH proposed a paradigm shift in machine perception, moving away from the "minimalist" reconstruction approach that characterized early computer vision. Instead of seeking the minimum information required for 3D reconstruction, they advocated for Animate Vision—a biological, behavior-driven model that exploits the information-rich nature of the real world.

The Challenge of Static Vision

Traditional computer vision often treated perception as a static, unidirectional pipeline: mapping an image to a 3D model. This approach proved brittle in dynamic environments. The authors identified a critical gap: biological vision is inherently active. Humans move their eyes to simplify computations and use behavior to define salience. The core challenge addressed is performing robust figure-ground segmentation in environments where the optimal visual cue may shift rapidly.

System Architecture and Methods

The proposed architecture rests on three functional pillars designed for real-time interaction:

  1. Visual Front-End (VFE): This layer computes "retinotopic" low-level features, such as directional derivatives at multiple scales, mimicking the early human visual cortex.
  2. Active Gaze Control: Utilizing the "Pahlavan" head-eye system, the robot performs fixation, target pursuit, and vergence to maintain focus on 3D targets.
  3. Multi-Cue Integration: The system processes three parallel feature maps—Motion Detection, Target Segmentation, and Disparity (Stereo)—to ensure robustness.
Figure 1: High-level system architecture for integrating multiple visual cues.

The architecture supports "Anytime Vision," a philosophy stating that a real-time system must provide an immediate, albeit coarse, response. Rather than waiting for complex initialization, the system refines its perception through continuous movement and data integration.

Experimental Evaluation

The system was evaluated using live video sequences at 25 Hz in a standard laboratory setting. The hardware consisted of a mobile robot equipped with the Pahlavan head. Unlike modern research relying on static datasets, this work focused on real-time interaction with humans in a dynamic environment.

Key Results and Insights

The primary demonstration of the system's efficacy was the "Distraction Test." In scenarios where a target is occluded by another moving object, systems relying solely on motion often switch to the more salient (closer) signal. By integrating disparity (depth) cues, the KTH system successfully ignored occluders by recognizing they were in a different depth plane than the target.

Figure 2: Real-time tracking performance during a distraction test.

Key insights from this work include:

  • The World as its Own Model: Active observers do not need perfect internal maps; they can sample the environment as needed.
  • Redundancy as Robustness: Parallel cues allow the system to maintain tracking even if one modality (e.g., color or motion) becomes unreliable.
  • Decoupling "Where" from "What": The system demonstrates that tracking can be achieved by clustering consistent motion and depth without requiring complex object recognition models.

Conclusion and Future Directions

The research proved that "rich" vision—integrating multiple complex cues—could be achieved at real-time speeds (25 Hz) even with 1990s hardware. While the system successfully utilized motion and depth, the authors noted that future work would involve more sophisticated cue selection logic and the addition of color and shape features to further enhance robustness.

Summary updated

BibTeX
@inproceedings{uhlin1995animate,
  author = {Tomas Uhlin and Jan-Olof Eklundh},
  title = {{Animate Vision in a Rich Environment}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
}
Figure