VR Image Depth Estimation Method Based on ResNeXt and Spatial Pyramid

Abstract

The rapid development of virtual reality technology has made depth estimation the key to enhancing immersion, but existing methods still suffer from insufficient multi-scale feature fusion and detail loss in complex scenes, leading to a decrease in depth map accuracy. To this end, a deep estimation network model based on Residual Networks with Next Generation (ResNeXt) and spatial pyramid modules is proposed. This model combines the efficient feature extraction of ResNeXt with the multi-scale fusion capability of Extremely Efficient Spatial Pyramid (EESP) module, combined with a hybrid attention mechanism, to improve depth estimation accuracy while optimizing computational efficiency. The experimental results show that the parameter count and floating-point operations of the proposed model in the training set are 19.2M and 33.2G, respectively, with an inference speed of 46FPS, demonstrating its robustness in indoor, outdoor, and low light environments. In addition, the model exhibits excellent performance in indoor, outdoor, and low light environments. In indoor scenes, the mean square error of the model is 0.045, the peak signal-to-noise ratio is 38.5, and the structural similarity is 0.92, which is 8% higher in accuracy than the baseline model. The results indicate that the method proposed by the research provides an effective and efficient solution for high-precision depth estimation in virtual reality applications