In this paper, we propose a deep learning-based video Compressive Sensing (CS) method for high-quality video recovery, that can be useful for diverse applications, like telemedicine and cloud-based surveillance. We split a video into a number of Groups Of Pictures (GOPs) with each GOP consisting of both keyframes and non-keyframes. The proposed video CS method uses a convolutional neural network (CNN) with Structural Similarity Index Measure (SSIM) based loss function. In a GOP, a convolution layer performs block-based frame-wise sampling, which optimizes the sample matrix. The recovery process has two stages. In the initial stage, CNN is employed to make efficient use of spatial redundancy. In the deep recovery stage, non-keyframes are compensated utilizing both keyframes and neighboring non-keyframes using multilevel feature compensation and single-level feature compensation, respectively. Through extensive experimentation, we establish the efficacy of our solution over several state-of-the-art image and video CS methods.