"Objective” methods to monitor physical activity and sedentary patterns in free-living conditions are necessary to further our understanding of their impacts on health. In recent years, many software solutions capable of automatically identifying activity types from portable accelerometry data have been developed, with promising results in controlled conditions, but virtually no reports on field tests. An automatic classification algorithm initially developed using laboratory-acquired data (59 subjects engaging in a set of 24 standardized activities) to discriminate between 8 activity classes (lying, slouching, sitting, standing, walking, running, and cycling) was applied to data collected in the field. Twenty volunteers equipped with a hip-worn triaxial accelerometer performed at their own pace an activity set that included, among others, activities such as walking the streets, running, cycling, and taking the bus. Performances of the laboratory-calibrated classification algorithm were compared with those of an alternative version of the same model including field-collected data in the learning set. Despite good results in laboratory conditions, the performances of the laboratory-calibrated algorithm (assessed by confusion matrices) decreased for several activities when applied to free-living data. Recalibrating the algorithm with data closer to real-life conditions and from an independent group of subjects proved useful, especially for the detection of sedentary behaviors while in transports, thereby improving the detection of overall sitting (sensitivity: laboratory model 24.9%; recalibrated model 95.7%). Automatic identification methods should be developed using data acquired in free-living conditions rather than data from standardized laboratory activity sets only, and their limits carefully tested before they are used in field studies.