CVPR Daily - Thursday

13 DAILY CVPR Thursday Learned Trajectory …. dynamic scene understanding and structure from motion for deformable or non-rigid objects. “This is just one instance of this huge field that is not yet solved 100%,” she points out. “With the tools available and the ideas in mind, it just seemed like a good fit at that moment!” The proposed method uses a neural network consisting of a feature extractor (encoder) and a subspace estimator (decoder). The feature extractor resembles PointNet, processing point trajectories independently and using 1D convolutional layers due to the temporal nature of the data. The subspace estimator includes regular MLPs that transform the features and incorporate a parametric model for subspace basis functions. “We combine this together to output the subspace models that correspond to the point trajectories,” Yaroslava explains. “We train the whole network as a network that tries to reconstruct trajectories as close to the original trajectories as possible, but also, we want it to be very good at clustering, so we incorporate the corresponding clustering loss, which is an InfoNCE loss.” One of the main difficulties she faced was the limited availability of data. Training a network that generalizes well to diverse scenarios is a challenge. To overcome this, she employed augmentations and parametric models to inject domain We observed that a sufficiently long trajectory can uniquely identify its corresponding motion model, which motivated our choice of the feature extractor construction. In theory, it is connected to the fact that motion models are low-dimensional subspaces in a high dimensional trajectory space, therefore different models intersect at zero only. UKRAINE CORNER

RkJQdWJsaXNoZXIy NTc3NzU=