Abstract
Shadow Stereo is a computer vision technique developed in 1987 to distinguish true object silhouettes from surface markings and detached shadows. By integrating multi-light source analysis with a simplified stereo check, the system categorizes edges based on their behavior under varying illumination and camera perspectives. This approach effectively isolates object boundaries on complex surfaces, treating shadows as informative geometric features rather than noise.
Introduction
In early computer vision, shadows were primarily viewed as obstacles or sources of error. Distinguishing between a dark surface marking and a shadow cast by an object is a non-trivial task for automated systems. Furthermore, "detached" shadows—those cast on the ground away from the object—can be easily confused with the object's actual boundary. This paper introduces "Shadow Stereo," a method designed to locate the precise silhouette of an object resting on a flat surface, even when the background contains significant visual noise or similar coloring.
Methodology
The Shadow Stereo approach utilizes two primary mechanisms to categorize image edges:
Multi-Light Source Analysis
The system captures multiple images from a fixed camera position while varying the light source's elevation. Crucially, the projected direction of light on the ground remains constant. This allows the system to differentiate between detached shadows (which move with the light) and attached shadows or surface markings (which remain stationary).
Simplified Stereo Check
A second camera view is used to determine if an edge lies on the ground plane or above it. By applying a geometric warp (homography) to align the two views based on the ground plane, the system can identify 3D structures. Edges that "move" after the warp are part of an object above the ground, while stationary edges are on the ground surface.
Edge Classification Logic
By combining these two checks, edges are classified into four distinct categories:
- LS-CS: Stationary under light change, stationary under camera change (Surface markings on the ground).
- LM-CS: Moves under light change, stationary under camera change (Detached shadows on the floor).
- LS-CM: Stationary under light change, moves under camera change (Attached object boundaries).
- LM-CM: Moves under both (Shadows cast onto other 3D objects).
Results and Discussion
The authors evaluated the method using a custom laboratory setup with a box-on-box scene. The supporting surface included arbitrary patterns to simulate industrial environments like conveyor belts.
The system successfully filtered out false positives from surface markings and detached shadows. Even when the object and background were visually similar, the oblique lighting provided enough geometric information to define the silhouette. The method proved robust against 49 different theoretical combinations of edge types.
Conclusion
Shadow Stereo demonstrates that shadows can be leveraged as a primary source of geometric data. By imposing constraints like a known ground plane and specific lighting configurations, the researchers transformed a complex 3D reconstruction problem into an efficient 2D filtering task. This approach was particularly valuable for the hardware constraints of the late 1980s, offering a computationally efficient path to reliable object segmentation.
Summary updated