Third-harmonic generation (THG) imaging of thick samples or large organisms requires TH light to be epicollected through the focusing objective. In this study we first estimate the amount of backward-to-forward TH radiation created by an isolated object as a function of size and spatial frequencies in the object. Theory and model experiments indicate that no significant signal can be epidetected from a (biological) dielectric structure embedded in a transparent medium. In contrast, backward emission is observed from metal nanoparticles where THG is partly a surface effect. We then address the case of an object embedded in a turbid medium. Experiments and Monte Carlo simulations show that epidetection is possible when the absorption mean free path of harmonic light in the medium exceeds its reduced scattering length, and that epicollection efficiency critically depends on the microscope field-of-view even at shallow depths, because backscattered light is essentially diffusive. These observations provide guidelines for optimizing epidetection in third-harmonic, second-harmonic, or CARS imaging of thick tissues.