The soft robots are welcomed in many robotic applications because of their high flexibility, which also poses a long-standing challenge on their proprioception, or measuring the real-time 3D shapes of the soft robots from internal sensors. The challenge exists in both the sensor design and robot modeling. In this paper, we propose a framework to measure the real-time high-resolution 3D shapes of soft robots. The framework is based on an embedded camera to capture the inside/outside patterns of the robots under different loading conditions, and a CNN to produce a latent code representing the robot state, which can then be used to reconstruct the 3D shape using a neural network improved from FoldingNet. We tested the framework on four different soft actuators with various kinds of deformations, and achieved real-time computation ($<$2ms/frame) for robust shape estimation of high precision ($<$5% relative error for 2025 points) at an arbitrary resolution. We believe the method could be widely applied to different designs of soft robots for proprioception, and enabling people to better control them under complicated environments. Our code is available at https://ai4ce.github.io/Deep-Soft-Prorioception/.