Access to the scene irradiance is a desirable feature in many computer vision algorithms. Applications like BRDF estimation, relighting or augmented reality need measurements of the object's photometric properties and the simplest method to get them is using a camera. However, the first step necessary to achieve this goal is the computation of the function that relates scene irradiance to image intensities. In this paper we propose to exploit the large variety of an object's appearances in photo collections to recover this non linear function for each of the cameras that acquired the available images. This process, also known as radiometric calibration, uses an unstructured set of images, to recover the camera's geometric calibration and a 3D scene model, using available methods. From this input, the camera response function is estimated for each image. This highly ill-posed problem is made tractable by using appropriate priors. The proposed approach is based on the empirical prior on camera response functions introduced by Grossberg and Nayar. Linear methods are proposed that allow to compute approximate solutions, which are then refined by non-linear least squares optimization.