Photorealistic Image Reconstruction from Hybrid Intensity and Event based Sensor
Event based sensors encode pixel wise contrast changes as positive or negative events at the instant they occur. This paradigm shift away from conventional image sensors provides advantages like low bandwidth requirement, low operating power requirement and low latency. However, in the process of encoding these contrast changes most of the spatial intensity information is lost. Previously there have been attempts at recovering this lost intensity information directly from the event data. Although the results are promising they lack the texture and the consistency of natural videos. We propose to reconstruct photorealistic intensity images by using a low frame rate conventional camera along with the event sensor. The low frame rate conventional camera provides us with texture-rich information, while the event sensor provides us with dense temporal measurements. We combine the strength of both by warping the low-frame rate video to intermediate locations where we only have event sensor data, resulting in high frame rate video. For accomplishing this task, we need the knowledge of scene depth and dense ego-motion estimates. We propose to use an unsupervised, learning free method to estimate depth from low-frame rate images. Unlike previous works, instead of estimating ego-motion directly from events, we propose an autoencoder based model to map events to pseudo-intensity frames. Using pseudo-intensity frames as input, we propose an unsupervised, learning free method adapted from Deep Image Prior to estimate the sensor ego-motion. Finally, we demonstrate photo-realistic reconstructions on a real hybrid sensor (DAVIS) dataset.
READ FULL TEXT