Scene analysis techniques in computer vision aim at obtaining a variety of information in a scene, such as a shape and visual textures of an object. For example, photometric stereo is a well-studied technique to obtain surface orientations of an object. However, photometric stereo cannot be applied for a translucent object because it assumes diffuse reflection on the surface but more complex optical phenomena occur in the translucent materials. Such assumptions in the scene analysis techniques have limited their opportunities of activity. A possible solution is a decomposition approach to extract a specific optical component. Since the optical phenomena are wavelength-dependent, spectral analysis could essentially improve the scene analysis techniques. In fact, analyzing optical components and a spectrum is important to obtain the visual textures. This thesis provides radiometric decomposition techniques toward obtaining a shape, an inner structure, and visual textures of an object. In general, detectors observe a radiometric intensity from the surface of a target object but the intensity includes various components, such as optical and thermal phenomena, different wavelengths, and different inner layers. By decomposing into each of the components, it can be easy to obtain the properties of the object. To achieve it, a unified framework to combine various techniques to extract an optical component is proposed and it enables a detailed decomposition. For spectral analysis, a novel one-shot hyperspectral imaging using faced reflectors is introduced. It is shown that multispectral images can be used for improving the robustness of photometric stereo. Moreover, analyzing far infrared light enables photometric stereo to work for transparent, translucent, and black materials. A novel pro-cam technique provides an ability to decompose inner layer images of an object, which let us understand the inner structure. Finally, a system which measures translucency, an important factor in visual textures, by a pro-cam system and physically reproduces it by a UV printer is implemented. Throughout this thesis, several abilities are achieved to segment image regions based on optical properties of materials and obtain spectral reflectance and a shape of an object whose materials are not only opaque but also even transparent, translucent, and black. Moreover, an inner structure of a translucent but non-scattering media can be obtained and a reproduction of translucency based on measurements is also achieved. Consequently, the proposed approach of radiometric decompositions is an effective solution to various scene analysis problems in computer vision.