Although highly diverse (Fonseca et al. 1996), the Atlantic forest mammal fauna is still poorly known, with very few sites exhaustively inventoried or subjected to long-term studies (Passamani et al. 2000). Although the first surveys using camera traps were carried out in the 1920s (e.g. Chapman 1927), most studies are rather recent (Karanth & Nichols 1998). This is not different in Brazil, where few studies have been published (Marques & Ramos 2001, Santos-Filho & Silva 2002, Silveira et al. 2003, Trolle 2003, Trolle & Kéry 2003). Given this, the objective of this paper is to assess the efficiency of camera trapping as an inventory technique for Neotropical forests in general and Atlantic forest in particular. The study was conducted at the Santa Lúcia Biological Station (SLBS), a biologically rich Atlantic Forest preserve located in south-eastern Brazil (Mendes & Padovan 2000) where mammals have been intensively live-trapped, observed from line-transects or had indirect evidence of their presence (faeces, footprints, scratches, etc.) recorded in earlier years (Passamani et al. 2000).

The distribution of species and population attributes are critical data for biodiversity conservation. As a tool for obtaining such data, camera traps have become increasingly common throughout the world. However, there are disagreements on how camera-trap records should be used due to imperfect species detectability and limitations regarding the use of capture rates as surrogates for abundance. We evaluated variations in the capture rates and community structures of mammals in camera-trap surveys using four different sampling designs. The camera traps were installed on internal roads (in the first and fourth years of the study), at 100-200 m from roads (internal edges; second year) and at 500 m from the nearest internal road (forest interior; third year). The mammal communities sampled in the internal edges and forest interior were similar to each other but differed significantly from those sampled on the roads. Furthermore, for most species, the number of records and the capture success varied widely among the four sampling designs. A further experiment showed that camera traps placed on the same tree trunk but facing in opposing directions also recorded few species in common. Our results demonstrated that presence or non-detection and capture rates vary among the different sampling designs. These differences resulted mostly from the habitat use and behavioral attributes of species in association with differences in sampling surveys, which resulted in differential detectability. We also recorded variations in the distribution of records per sampling point and at the same spot, evidencing the stochasticity associated with the camera-trap location and orientation. These findings reinforce that for species whose specimens cannot be individually identified, the capture rates should be best used as inputs for presence and detection analyses and for behavior inferences (regarding the preferential use of habitats and activity patterns, for example). Comparisons between capture rates or among relative abundance indices, even for the same species, should be made cautiously.