Abstract: Particle Picking in cellular cryo-electron tomograms (cryo-ET) is crucial for \textit{in situ} structure detection of macromolecules and protein complexes. Given the problems associated with the traditional template-matching approaches for particle picking, learning-based solutions are necessary for particle picking. A big challenge in this regard is the lack of annotated data for training. In this work, we present TomoPicker, a Positive-Unlabeled learning-based annotation-efficient particle-picking approach that can effectively pick particles when only a minuscule portion ($\sim 0.3-0.5\%$) of the total particles in a cellular cryo-ET dataset are provided for training. We evaluated our method on a benchmark cryo-ET dataset of eukaryote cells, where we observed about $30\%$ improvement by TomoPicker against the most recent state-of-the-art annotation efficient learning-based picking approaches.