A novel approach to pyramid training and feature size reduction

In this study, we address the challenge of high GPU memory requirements in state-of-the-art stereo image matching models. Given limited computational resources, we propose a novel technique called Pyramid Training Approach with Similarity Comparison to reduce the size requirements of machine learning models. Our method involves training small neural networks independently and subsequently combining them into larger networks. Once the training of an individual network is completed, it is combined with another network to form a new, larger network, which is then further trained. This iterative process continues, progressively building larger networks. During this process, the weights of the initial networks can be frozen once they reach sufficient training maturity, thereby conserving memory. Additionally, we employ a similarity comparison between features in the output feature maps during the combination of different networks to identify and merge redundant features using a similarity matrix. This approach effectively reduces the size of the feature maps, thus lowering the overall memory footprint. Our experimental results demonstrate that this method significantly reduces the GPU memory requirements while maintaining competitive performance in different domains. This technique offers a viable solution for deploying large-scale models on limited hardware resources, facilitating broader accessibility and application.