Abstract: We present an approach for building an affine representation of an unknown curved object viewed under orthographic projection from images of its occluding contour. It is based on the observation that the projection of a point on a curved, featureless surface can be computed along a special viewing direction that does not belong to the point's tangent plane. We show that by circumnavigating the object on the tangent plane of selected surface points, we can (1) compute two orthogonal projections of every point projecting to the occluding contour during this motion, and (2) compute the affine coordinates of these points. Our approach demonstrates that affine shape of curved objects can be computed directly, i.e., without Euclidean calibration or image velocity and acceleration measurements.
The Challenge of Smooth Surfaces
In computer vision, "shape from silhouette" is a classic challenge. For objects with sharp corners or distinct markings, tracking points across frames is straightforward. However, for smooth, curved objects, the silhouette (or occluding contour) is generated by the visible rim—the set of points where the line of sight is perfectly tangent to the surface.
The primary difficulty is that as the camera moves, the visible rim "slides" over the surface. Consequently, the observer is not looking at the same physical points from one frame to the next. Traditional solutions attempt to model this sliding motion using differential geometry, measuring image velocities and accelerations. However, these methods are highly sensitive to noise and typically require a perfectly calibrated camera. This research seeks a reliable 3D representation—specifically an affine reconstruction—without relying on these fragile measurements.
Purposive Viewpoint Control
The paper proposes an Active Affine Shape Recovery Strategy. Instead of passively observing an object, the observer "purposively" controls its motion to find specific geometric alignments. This approach trades complex, noise-sensitive calculations for controlled movement.
Key Methodological Pillars:
- Planar Motion & Epipolar Geometry: The camera moves within a single plane. This simplifies the 3D reconstruction problem into a series of 2D slices or epipolar planes.
- The Extremum Trick: The author demonstrates that if you circumnavigate an object, the distance between the silhouette and a reference point reaches a local maximum or minimum exactly when the camera is aligned with the "tangential" or "normal" directions of a surface point.
- Virtual Points: To create a coordinate system without physical markers, the paper introduces "virtual points." These are 3D points defined by the intersection of visual rays from two different views, acting as stable spatial anchors to establish an affine reference frame.
Experimental Validation
The approach was evaluated using both simulated and real-world data:
- Simulated Data: Synthetic elliptical cross-sections were used to test sensitivity to noise. Even with image noise at 0.8% of the object's size, an iterative repositioning method for virtual points kept reconstruction errors low.
- Real-World Data: A "cup" experiment was conducted using a PUMA 700 robot arm to rotate a featureless cup. Four markers on the platform provided the initial affine basis. The system successfully recovered the circular cross-section of the cup, demonstrating that the method can localize the "center of the osculating circle" (the center of local curvature) for various surface points.
Conclusion and Future Work
This research proves that affine shape can be computed directly by choosing where to look, making the underlying mathematics simpler and more robust. While the current model assumes orthographic projection and is restricted to planar motion, it provides a foundation for richer geometric descriptions of featureless objects. Future work aims to extend these principles to perspective projection and fully automated "hand-eye" robotic systems.
Summary updated