At present, the research of object counting mostly focuses on single object counting, such as crowd counting and vehicle counting. In real life scenarios, it is common that crowds and vehicles appear at the same time, so multi-class object counting is more valuable. However, multi-class object counting has not been drawing broad attentiion yet. This study first provides a dataset of real scene multi-class counting (RSMCOC) for the research of multi-class object counting, and then proposes a lightweight multi-class counting network (LMCNet). In LMCNet, we design a Ghost attention module, adopting the Ghost attention mechanism that is a hybird mechanisem of channel and spatial attention mechanisms. To obtain high quality multi-channel density maps with low computing efficiency, the Ghost attention module takes into account the phenomenon of ''feature map distortion" and mixes the inherent and Ghost feature maps, where the latter is generated by a Ghost module. Moreover, LMCNet develops a novel loss function, called Focal-L2, which is suitable for multi-class object counting tasks. This loss function can adaptively allocate weights to the loss of each category, avoiding the problem caused by class imbalance. For comparison, we extend existing single object counting algorithms to multi-class counting tasks by changing the output channels of these algorithms. Through extensive experiments on the RSMCOC dataset, we demonstrate that the proposed model achieves a better balance of counting performance and computing efficiency compared to existing models.