CS Semineri: “Efficient Multi-Agent Trajectory Prediction and Unknown Object Segmentation”, Fatma Güney, 14:00 18 Nisan (EN)

“Efficient Multi-Agent Trajectory Prediction and Unknown Object Segmentation”

Asst. Prof. Fatma Güney
Department of Computer Engineering, Koç University

Date-Time: Tuesday, April 18, 2023 at 14:00 p.m.
Place : EA-409

Abstract:
In the first part, we will assume perfect perception information and talk about forecasting the future trajectories of agents in complex traffic scenes. While safe navigation requires reliable and efficient predictions for all agents in the scene, the standard paradigm is simply predicting the trajectory of a single agent of interest by orienting the scene according to the agent. The straightforward extensions to multi-agent predictions result in inefficient solutions that require iterating over each agent. We propose ADAPT, a novel approach for jointly predicting the trajectories of all agents in the scene with dynamic weight learning. We demonstrate that the prediction head can adapt to each agent’s situation efficiently. ADAPT outperforms the state-of-the-art methods at a fraction of their computational overhead on both single-agent setting of the Argoverse and multi-agent setting of the Interaction. Our analyses show that ADAPT can focus on each agent with adaptive prediction without sacrificing efficiency.

In the second part, we will break that assumption and focus on unknown object segmentation. Semantic segmentation methods typically perform per-pixel classification by assuming a fixed set of semantic categories. While they perform well on the known set, the network fails to learn the concept of objectness, which is necessary for identifying unknown objects. We explore the potential of query-based mask classification for unknown object segmentation. We discover that object queries specialize in predicting a certain class and behave like one vs. all classifiers,
allowing us to detect unknowns by finding regions that are ignored by all the queries. Based on a detailed analysis of the model’s behavior, we propose a novel anomaly scoring function. We demonstrate that mask classification helps to preserve the objectness and the proposed scoring function eliminates irrelevant sources of uncertainty. Our method achieves consistent improvements in multiple benchmarks, even under high domain shift, without retraining or using outlier data. With modest supervision for outliers, we show that further improvements can be achieved without affecting the closed-set performance.

Bio:
Fatma Güney is an assistant professor at the Dept. of Computer Engineering at Koç University in Istanbul. She received her Ph.D. from MPI in Germany and worked as a postdoctoral researcher at VGG in Oxford. Currently, she is leading a small team called Autonomous Vision Group (AVG) at the KUIS AI center. They work on a range of topics including but not limited to monocular depth estimation and semantic segmentation, multi-object tracking, unsupervised video object segmentation, end-to-end learning of driving, and stochastic future prediction.