Weakly Supervised Bird's Eye View Motion Prediction With Object Prior

Abstract

Current motion prediction methods mainly focus on predicting the motion information for each grid cell on the bird’s eye view representation. However, they predict motion for every single cell without object-level constraints, thereby suffering from limited performance. Inspired by the observation that the scenes can be segmented into a collection of objects with similar motion patterns, we propose an object-aware motion prediction framework to encode the object-level motion information into both the feature and motion spaces for more consistent motion results. Specifically, we design a probabilistic motion feature encoding module to generate probabilistic object-aware features. It utilizes the object priors in the current and future frames to parameterize two probability distributions, i.e., the current distribution and future distribution. Since the future frames are invisible during inference, we introduce a Kullback-Leibler divergence loss to minimize the divergence between these two distributions. By encoding the sampled latent code from the distributions into motion feature space, we can obtain more accurate motion results. Furthermore, we propose a confidence-aware motion encoding module to explicitly aggregate the object-level motion with learnable confidence weights, which aims to preserve the consistent motion patterns of points in the same objects. We conduct thorough experiments on the nuScenes and Waymo datasets. Specifically, our method achieves the mean errors of 0.2670 meters and 0.2807 meters on the slow groups of the two benchmarks, respectively, outperforming the compared weakly supervised method by up to 17%.