Inexpensive and commercial cameras can often handle a fixed bandwidth, often leading to a tradeoff between the desired signal's angular, spatial, and temporal resolution. This talk will discuss techniques to overcome this constraint and acquire videos with (a) high temporal resolution and (b) high angular resolution.

High frame-rate videos or high temporal resolution videos find applications in science, industry, and entertainment. However, the frame rate of most cameras is fixed at 30 frames per second. Several coded exposure techniques have been proposed to compressively acquire high-frame-rate videos at low-bandwidth, avoiding the cost of higher-frame-rate cameras. Recently, a Coded-2-Bucket camera has been proposed that captures two compressed measurements in a single exposure, while other techniques capture a single measurement. However, the quantitative and qualitative advantages of two measurements over a single one are yet unknown. Here, we propose a unified learning-based framework to make such a qualitative and quantitative comparison between those that capture only a single coded image and capture two measurements per exposure (C2B). Our learning-based framework consists of a shift-variant convolutional layer followed by a fully convolutional deep neural network. While achieving state-of-the-art reconstruction in all three techniques, we can further compare each of their performances. The C2B sensor has a significant advantage over acquiring a single pixel-wise coded measurement. However, when most scene points undergo motion, the C2B sensor has only a marginal benefit over the single pixel-wise coded exposure measurement.

Next, we look at acquiring videos with a high angular resolution, also known as light-field videos. Acquiring LF data with high angular, spatial and temporal resolution poses significant challenges, especially with space constraints preventing bulky optics. We explore the application of small baseline stereo videos for reconstructing high-fidelity LF videos. We propose a self-supervised learning-based algorithm for LF video reconstruction from stereo video. We guide the reconstruction via the geometric information from the individual stereo pairs and the temporal information from the video sequence. A low-rank constraint based on layered LF displays is used to regularize the reconstruction further. Quantitatively the LF videos show higher fidelity than previously proposed unsupervised approaches for LF reconstruction. We demonstrate our results via LF videos generated from publicly available stereo videos acquired from commercially available stereoscopic cameras. Finally, we show that our reconstructed LF videos allow applications such as post-capture focus control and RoI-based focus tracking for videos.