Generalization vs. Personalization: A Dynamic Layer Masking Solution for Data Heterogeneous Federated 3D Object Detection
J.T. Verhoog (TU Delft - Mechanical Engineering)
A.B. Ünal – Mentor (TU Delft - Mechanical Engineering)
Holger Caesar – Mentor (TU Delft - Mechanical Engineering)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Robust perception in intelligent vehicles demands 3D object detectors that remain dependable under domain shifts, such as changes in time of day, location, or weather.
These factors result in heterogeneous data, where domain-specialized models may be advantageous.
Naively retraining models for each scenario is computationally expensive, energy-intensive, and possibly infeasible due to data scarcity.
Instead, federated learning offers an efficient alternative to costly retraining for multi-domain adaptation.
This study explores how federated strategies for addressing data heterogeneity can improve the cross-domain robustness of 3D object detectors to temporal variations.
Traditionally, these strategies require manually selecting which parts of the model are globally aggregated and which are locally personalized.
To resolve this manual choice, a new Centered Kernel Alignment (CKA)-based strategy, FedCKA, is proposed.
It dynamically handles the personalization-generalization trade-off by selectively sharing representation-consistent layers across clients.
Evaluation on a unified multi-domain benchmark on nuScenes shows that FedCKA outperforms established federated baselines, including FedBN, FedRep, and FedSelect, in heterogeneous environments.
The findings offer both a comparative benchmark and a promising direction for federated, location-, weather-, and illumination-robust 3D perception.