This paper examines how camera motion provides information about the shape of an object when its reflectance properties (BRDF) are unknown. Motivated by psychophysical evidence that motion cues convey shape information independently of material properties, we derive a differential stereo relation that relates camera motion to the depth of a surface with unknown isotropic BRDF, which generalizes traditional Lambertian assumptions. Under orthographic projection, we show that two camera motions can produce shape invariants for several restricted material classes, although shape constraints may not exist in general. Under perspective projection, we prove that three differential camera motions suffice to recover surface depth along with unknown isotropic BRDF and directional lighting parameters.

Recovering the 3D shape of objects from images is a fundamental goal of computer vision. Classical shape-from-shading and multi-view stereo approaches typically assume known or Lambertian reflectance, which limits their applicability to real-world objects with complex materials. In reality, most surfaces exhibit non-Lambertian reflectance — metals are specular, plastics show mixed diffuse-specular behavior, and many materials have complex BRDFs (Bidirectional Reflectance Distribution Functions). The question we address is: what can we infer about shape from camera motion alone, when the BRDF is completely unknown? This question has both practical significance for robust 3D reconstruction and theoretical interest in understanding the fundamental limits of shape recovery.

Prior work on shape recovery under unknown reflectance falls into several categories. Photometric stereo methods vary lighting conditions rather than camera position, and extensions to unknown BRDF require multiple lights or special configurations. Shape from motion approaches (e.g., structure from motion) rely on feature correspondences and are independent of reflectance but require textured surfaces. Helmholtz stereopsis exploits BRDF reciprocity to handle arbitrary reflectance but requires co-located light-camera pairs. Our work is unique in that we derive shape constraints purely from differential camera motion under fixed illumination, without requiring special lighting setups, feature correspondences, or BRDF knowledge.

We derive a differential stereo relation that connects infinitesimal camera motion to surface depth for an object with unknown isotropic BRDF. Consider a surface point observed under fixed directional illumination. As the camera undergoes an infinitesimal motion, the observed intensity change depends on both the surface geometry (through the change in viewing direction) and the BRDF (through the reflectance gradient). The key insight is that by considering ratios of intensity changes under different camera motions, we can eliminate the BRDF dependence and obtain constraints that involve only the surface depth. This generalizes the classical Lambertian differential stereo to arbitrary isotropic BRDFs.

Under orthographic projection, we analyze what shape information can be recovered from camera motion with unknown isotropic BRDF. We prove that two camera motions can yield shape invariants for several restricted BRDF classes, including Lambertian, specular, and certain parametric models. However, we also establish a fundamental negative result: for a completely general isotropic BRDF under orthographic projection, shape constraints from camera motion may not exist. This theoretical limit is inherent to the problem itself, not an artifact of any specific reconstruction algorithm. The negative result motivates moving to the perspective projection model, where additional geometric information becomes available.

The perspective projection model provides additional geometric constraints compared to the orthographic case. We prove the key positive result: three differential camera motions under perspective projection suffice to recover the surface depth, along with the unknown isotropic BRDF and directional lighting parameters. The intuition is that perspective projection creates a depth-dependent relationship between image position and viewing direction, which provides additional equations that break the ambiguity present in the orthographic case. We derive closed-form expressions for the depth in terms of the observed intensity changes and demonstrate that the solution is unique up to a global scale ambiguity, which is inherent to monocular 3D reconstruction.

We validate our theoretical results through both synthetic and real experiments. Synthetic experiments on objects with known ground-truth geometry and various BRDFs (Lambertian, Phong, Cook-Torrance, measured BRDFs) confirm that the derived differential stereo relations yield accurate depth recovery. We demonstrate robustness to moderate noise levels in the synthetic setting. Real experiments are conducted with a controlled setup using a turntable and fixed illumination, photographing objects with diverse materials including ceramic, plastic, and metallic surfaces. The reconstructed shapes qualitatively match the true object geometry, validating that the theory translates from the differential limit to practical finite camera motions.

We have presented a theoretical analysis of what camera motion reveals about shape when the surface BRDF is unknown. Our contributions include: a differential stereo framework generalizing Lambertian assumptions, impossibility results under orthographic projection, and a constructive result showing that three perspective camera motions suffice for complete depth recovery with unknown isotropic BRDF. These results characterize fundamental limits of shape recovery that are inherent to the problem, independent of any particular algorithm. Future work may explore extensions to anisotropic BRDFs, spatially varying materials, and unknown lighting configurations.