Recently, data-driven feed-forward 3D reconstruction methods, such as DUSt3R, MASt3R, Fast3R and VGGT, have gained widespread attention due to their superior end-to-end processing capabilities across various geometric 3D vision tasks. However, heavy reliance on GPU hardwares limits the applicability of these 3R methods to only single-image pairs or small-scale datasets, making them challenging to handle large-scale high-resolution photogrammetric images. In this work, we conduct a survey on these 3R methods and employ a divide-and-conquer framework that divides the entire image dataset into several overlapping sub-blocks, reconstructs each sub-block separately using 3R methods, and then merges them per 3D similarity transformations. Experimental results demonstrate that our method effectively expands the number of images that the aforementioned feed-forward 3R methods can handle. Furthermore, a comprehensive experiment on photogrammetric data is carried out by comparing the processing time, GPU memory usage, and accuracy to explore the possibility of applying these novel feed-forward 3R methods to high-resolution photogrammetric datasets. Project web: <code>https://sh1nzzz.github.io/3R-methods-via-divide-and-conquer-strategy.github.io/</code>.