This focus paper provides an overview on the emerging area of in-memory neural vision chips for building intelligent cameras. The generalization abilities of neural networks and their close proximity to eye and sensory processing units in human brain lays foundations for building neural vision chips capable of performing object detection, classification, and predictions. In real-time settings, often the information visible to a camera does not meaningfully capture all the relevant context, object dynamics, and behaviors, making smart camera solutions complex and energy hungry. Learning from images is a continuous process in the human brain while consuming extremely low energy, and these properties have proven to be the benchmarks of the ideal neural vision chips. Offloading the intelligent data processing and image processing to edge devices using neural vision chips offers several advantages of speed, lower energy costs, and security. In practice, with emerging memory arrays, memory and computing become indistinguishable that can perform dot-product computations, store weights, and perform logic operations forming the basic blocks of building several types of neural networks useful for solving a range of computer vision problems.