Processor and method for simultaneous localization and mapping based on real-time neural network rendering using sparse mixture-of-experts model acceleration architecture

A method for SLAM based on real-time neural network rendering using a SLAM processor configured to perform SLAM based on real-time neural network rendering, the method comprising:

• a sampling step of hierarchically sampling, by the SLAM processor, a 2D image collected through any image collection device and pose information of the image collection device corresponding to the 2D image;
• and
• a rendering step of performing, by the SLAM processor, SLAM through real-time rendering for data sampled in the sampling step,
• wherein the rendering step comprises:
• a scheduling step of scheduling a processing order of input batches to improve computational efficiency;
• and
• a computation step of performing a real-time neural network operation based on a sparse expert model including a plurality of expert neural networks and reducing a number of neural network channels by exclusively activating an expert neural network differently selected for each input batch, and performing a neural network operation on the input batches based on information scheduled in the scheduling step.