The emerging intelligent reflecting surface (IRS) can significantly improve the system capacity, and it has been regarded as a promising technology for the beyond fifth-generation (B5G) communications. For IRS-assisted multiple input multiple output (MIMO) systems, accurate channel estimation is a critical challenge. This severely restricts practical applications, particularly for resource-limited indoor scenario as it contains numerous scatterers and parameters to be estimated, while the number of pilots is limited. Prior art tackles these issues and associated optimization using mathematical-based statistical approaches, but are difficult to solve as the number of scatterers increase. To estimate the indoor channels with an affordable piloting overhead, we propose an offset learning (OL)-based neural network for channel estimation. The proposed OL-based estimator can dynamically trace the channel state information (CSI) without any prior knowledge of the IRS-assisted channel structure as well as indoor statistics. In addition, inspired by the powerful learning capability of convolutional neural network (CNN), CNN-based inversion blocks are developed in the offset estimation module to build the offset estimation operator. Numerical results show that the proposed OL-based estimator can achieve more accurate indoor CSI with a lower complexity as compared to the benchmark schemes.