Abstract
This paper introduces the Image Torque Operator, a bottom-up tool inspired by classical physics to identify the closure of contours in mid-level vision. By treating image edges as physical forces and calculating the torque relative to a center point, the operator provides a robust, scale-invariant measure of "object-ness." Evaluation on standard benchmarks shows significant improvements in edge detection, saliency mapping, and image segmentation.
Introduction
In computer vision, researchers have achieved high performance in low-level edge detection and high-level object recognition. However, the "mid-level" stage—where local edges are grouped into meaningful structures like boundaries and shapes—remains a complex challenge. Traditional edge detectors often struggle to distinguish meaningful object boundaries from background texture. This research introduces a mathematical operator that naturally favors the Gestalt principle of closure, identifying closed loops without requiring massive training datasets.
Methods
The core innovation is the Torque Operator, which borrows directly from the physics of rotational force: $\vec{\tau} = \vec{r} \times \vec{F}$.
In this framework, an image edge is treated as a unit force vector ($\vec{F}$) acting along its tangent. The displacement vector ($\vec{r}$) is the path from the center of a local image patch to that edge point. For every pixel in a patch, the operator calculates the torque. If edges are randomly oriented (texture), their torque values cancel out. If edges are aligned to form a circle or convex shape around the center, the torque values sum to a large magnitude. This provides a geometric link where the torque of a closed curve is proportional to the area it encloses.
To handle objects of varying sizes, the operator is applied across multiple patch sizes, creating a "Torque Volume." From this, a Torque Value Map (the strongest response across scales) and a Scale Map (the size of the patch producing that response) are extracted. The computation is made efficient ($O(N)$ time) using integral images, making it practical for real-time applications.
Results
The Torque Operator was validated across several standard benchmarks, including the Berkeley Segmentation Dataset (BSDS500), the Judd Eye-Tracking Dataset, and Caltech-101.
Key results include:
- Boundary Detection: Adding torque information to standard methods like Canny or gPb consistently improved F-measure scores. On the Berkeley benchmark, it boosted Canny-based detection from 0.20 to 0.23.
- Segmentation: In multi-region segmentation, the torque-based approach achieved a covering score of 0.429, significantly outperforming the classic Normalized Cut (0.390).
- Object Recognition: When combined with SIFT descriptors on Caltech-101, the "Multiscale Torque" (MST) descriptor provided a 2.45% to 4.66% accuracy gain over SIFT alone, proving that geometric contour information complements texture-based features.
- Texture Suppression and Saliency: Because the operator relies on edge polarity, random textures cancel out while boundaries provide consistent rotational "push." The operator produces extrema at the center of objects, making it an effective tool for visual attention and fixation.
Conclusion
The Image Torque Operator acts as a bridge between local edges and global shapes by leveraging physical principles. While currently optimized for convex shapes, future work could generalize the operator to handle more complex topologies or extend it to the spatio-temporal domain to group moving edges into coherent objects.
Summary updated