|
Removing dense foreground occlusion from images and reconstructing the target of interest is a critical vision task. In previous studies, it was generally tackled through frame-based methods, but the performance was limited due to the lack of valid information. With the development of event cameras, their advantages in high temporal resolution and asynchronous response mechanism at each pixel have shown significant potential in various visual tasks. However, the event stream is plagued by multiple noise factors and static perceptual limitations, making it difficult to directly restore the local texture and absolute color of occluded objects. To overcome these challenges, we incorporate event stream information into the image frame restoration process to achieve a more effective occlusion removal. Specifically, we introduce a hybrid neural network for removing foreground occlusions from event-frame inputs, along with the design of an event stream encoder based on Spiking Neural Networks (SNN) and a Temporal Channel Attention Block (TCA) to enhance frame features. In addition, in order to significantly enhance the capability of occlusion removal, we introduce a General Restoration Block (GRB), which is applicable to both event data and frame data. Extensive experimental results indicate that the proposed method performs favorably against the state-of-the-art approaches.
|
