Multi-camera three-dimensional (3D) pose estimation (MCTPE) has already achieved very high estimation accuracy by utilizing deep neural network (DNN) based models. However, long inference latency of the utilized complex DNN models prevents the real-time deployment of MCTPE. Device-edge collaborative inference (co-inference) is a promising way to reduce the total inference latency of MCTPE, which performs one part of the inference operations on the devices and the other part of inference operations on the edge server to fully exploit computation resources of both the devices and the edge server. Besides, there is overlap between the detection ranges of different cameras in many cases. We propose the camera-selecting device-edge collaborative inference for MCTPE (CDC-MCTPE), which discards some of the raw data from parts of the cameras to reduce the inference task size without sacrificing estimation accuracy too much. In CDC-MCTPE, we formulate the joint optimization problem with regard to the model split points and camera-selecting decisions to minimize the total inference latency and the energy consumption of all devices under the constraints of the estimation accuracy. A Random-Ordered Greedy Algorithm (ROGA) is proposed to quickly solve the problem. The simulation results show that the proposed CDC-MCTPE achieves better performance compared with three benchmarks.