Adobe, Inc.

Seoungwug Oh

Joon-Young Lee

Kalyan Krishna Sunkavalli

Ning Xu

Certain aspects involve using a space-time memory network to locate one or more target objects in video content for segmentation or other object classification. In one example, a video editor generates a query key map and a query value map by applying a space-time memory network to features of a query frame from video content. The video editor retrieves a memory key map and a memory value map that are computed, with the space-time memory network, from a set of memory frames from the video content. The video editor computes memory weights by applying a similarity function to the memory key map and the query key map. The video editor classifies content in the query frame as depicting the target feature using a weighted summation that includes the memory weights applied to memory locations in the memory value map.

Space-time memory network for locating target object in video content

Various embodiments describe video object segmentation using a neural network and the training of the neural network. The neural network both detects a target object in the current frame based on a reference frame and a reference mask that define the target object and propagates the segmentation mask of the target object for a previous frame to the current frame to generate a segmentation mask for the current frame. In some embodiments, the neural network is pre-trained using synthetically generated static training images and is then fine-tuned using training videos.

Video object segmentation by reference-guided mask propagation

In implementations of segmenting objects in video sequences, user annotations designate an object in any image frame of a video sequence, without requiring user annotations for all image frames. An interaction network generates a mask for an object in an image frame annotated by a user, and is coupled both internally and externally to a propagation network that propagates the mask to other image frames of the video sequence. Feature maps are aggregated for each round of user annotations and couple the interaction network and the propagation network internally. The interaction network and the propagation network are trained jointly using synthetic annotations in a multi-round training scenario, in which weights of the interaction network and the propagation network are adjusted after multiple synthetic annotations are processed, resulting in a trained object segmentation system that can reliably generate realistic object masks.

Growing community of inventors

Seoul, South Korea

Seoungwug Oh