AI-based Holl-filling with Temporal Consistency

Depth image-based rendering (DIBR) has been recently considered as a significant technology for generating 3D virtual views. However, the hole-filling method is still the main challenge in the DIBR engine because of the occlusions of foreground objects. While looking into the video sequences, the occluded parts in a frame may be revealed in the past or future frames. The temporal information could be the useful cues to recover the occlusions of the current frame. In this paper, we design an encoder-decoder neural model with a bounded region attention module to effectively fill the holes. This attention module is aim to extract the useful hints from neighbor frames.