Computer Vision News - July 2020

Oral Presentation 56 standard learning methods. As a consequence, exploiting global graph information for MOT has remained an unexplored territory, and most work has been devoted to pairwise (edge) feature extraction . In recent years, though, a new family of deep learning models that can operate directly on graph-structured data, known as message passing neural networks, has been introduced. In our work, we show how these models can be used to learn to solve the data association problem and replace expensive combinatorial optimization solvers. Give me more details! We propose a model that takes a set of RGB frames and object detections as input, and leverages both appearance and scene geometry cues over a graph to learn to predict the links among object detections forming trajectories. The key component of our model is a message passing network that can reason about a set of object detections jointly and encode high-order interactions among them. Previous methods had to introduce handcrafted terms and deal with complex "A fully differentiable model for data association that can replace expensive optimization algorithms while running at a fraction of their cost and yielding significantly better results" optimization frameworks in order to incorporate such interactions. Instead, our model can learn them implicitly, via neural message passing. As a result, our method yields significantly improved tracking results and is particularly good at identity preservation. Not only that, but our model only requires a forward pass through a convolutional network and a set of lightweight multi-layer perceptrons plus a simple rounding scheme . Hence it is much faster than Best of CVPR 2020