We propose a learning-based method for detecting carried objects that generates candidate image regions from protrusion, color contrast and occlusion boundary cues, and uses a classifier to filter out the regions unlikely to be carried objects. The method achieves higher accuracy than state of the art, which can only detect protrusions from the human shape, and the discriminative model it builds for the silhouette context-based region features generalizes well. To reduce annotation effort, we investigate training the model in a Multiple Instance Learning framework where the only available supervision is "walk" and "carry" labels associated with intervals of human tracks, i.e., the spatial extent of carried objects is not annotated. We present an extension to the miSVM algorithm that uses knowledge of the fraction of positive instances in positive bags and that scales to training sets of hundreds of thousands of instances.