The field of 3D reconstruction and perception is witnessing a significant paradigm shift with the advent of deep learning-based methods. These techniques are revolutionizing the way 3D structures are recovered from 2D images, enabling applications such as augmented reality, autonomous driving, and robotics. Recent developments have focused on improving the accuracy, scalability, and robustness of these models, particularly in challenging scenarios like texture-less regions and dynamic scenes. Notably, feed-forward approaches have emerged as a promising direction, allowing for joint inference of camera poses and dense geometry in a single forward pass. Furthermore, innovative methods for temporal fusion, multimodal perception, and relative positional encoding are being explored to enhance the performance of 3D perception tasks. Noteworthy papers in this area include OnlineBEV, which achieves state-of-the-art performance in camera-only 3D object detection through recurrent temporal fusion. MonoMVSNet is another notable work, which integrates monocular priors into multi-view stereo networks to improve depth estimation in challenging regions. Additionally, Cameras as Relative Positional Encoding presents a novel relative encoding technique that captures complete camera frustums, leading to improved performance in feed-forward novel view synthesis and other tasks.