NexusFlow: Unifying Disparate Tasks under Partial Supervision via Invertible Flow Networks

Fangzhou Lin, Yuping Wang, Yuliang Guo, Zixun Huang, Xinyu Huang, Ziming Zhang, Zhengzhong Tu, Haichong Zhang, Kazunori Yamada, Liu Ren

Keywords: Partially Supervised Multi-Task Learning ; Autonomous Driving;

Abstract: Partially Supervised Multi-Task Learning (PS-MTL) aims to leverage knowledge across tasks when annotations are incomplete. Existing approaches, however, have largely focused on the simpler setting of homogeneous, dense prediction tasks, leaving the more realistic challenge of learning from structurally diverse tasks unexplored. This paper addresses this critical gap by introducing NexusFlow, a novel, lightweight, and plug-and-play framework. We establish a challenging new benchmark where supervision for the highly disparate tasks of dense map reconstruction and sparse multi-object tracking is split across different geographic domains, compounding task heterogeneity with a significant domain gap. NexusFlow introduces a pair of surrogate networks with invertible coupling layers to align the latent feature distributions of these tasks, creating a unified representation that enables effective knowledge transfer. We validate our framework's effectiveness on these core perception tasks in autonomous driving, demonstrating state-of-the-art results on the nuScenes benchmark. Our approach significantly outperforms strong partially supervised baselines. Our code and video demos are available in the supplementary material.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 3901