Navigating Data Heterogeneity in Federated Learning A Semi-Supervised Federated Object Detection

TL;DR

This work tackles data heterogeneity in federated learning for object detection by introducing Semi-Supervised Federated Object Detection (SSFOD) and the two-stage FedSTO framework. It combines selective training of the backbone, full-parameter training with orthogonal regularization, and local EMA-driven pseudo labeling to leverage unlabeled client data while preserving privacy. validated on BDD100K, Cityscapes, and SODA10M, FedSTO outperforms existing federated and semi-supervised baselines and approaches fully supervised centralized performance with only ~20-30% labeled data, while reducing communication bandwidth. The approach advances practical privacy-preserving learning under domain shifts common in autonomous driving and offers design guidance for handling data heterogeneity in SSFOD.

Abstract

Federated Learning (FL) has emerged as a potent framework for training models across distributed data sources while maintaining data privacy. Nevertheless, it faces challenges with limited high-quality labels and non-IID client data, particularly in applications like autonomous driving. To address these hurdles, we navigate the uncharted waters of Semi-Supervised Federated Object Detection (SSFOD). We present a pioneering SSFOD framework, designed for scenarios where labeled data reside only at the server while clients possess unlabeled data. Notably, our method represents the inaugural implementation of SSFOD for clients with 0% labeled non-IID data, a stark contrast to previous studies that maintain some subset of labels at each client. We propose FedSTO, a two-stage strategy encompassing Selective Training followed by Orthogonally enhanced full-parameter training, to effectively address data shift (e.g. weather conditions) between server and clients. Our contributions include selectively refining the backbone of the detector to avert overfitting, orthogonality regularization to boost representation divergence, and local EMA-driven pseudo label assignment to yield high-quality pseudo labels. Extensive validation on prominent autonomous driving datasets (BDD100K, Cityscapes, and SODA10M) attests to the efficacy of our approach, demonstrating state-of-the-art results. Remarkably, FedSTO, using just 20-30% of labels, performs nearly as well as fully-supervised centralized training methods.

Figures (5)

• Figure 1: An overview of our FedSTO method within the SSFOD framework with key components: selective training, orthogonal enhancement, and local Exponential Moving Average (EMA)-driven pseudo label assignment, organized into two stages. Algorithm steps are numbered accordingly.
• Figure 2: Performance comparison on BDD100K dataset yu2020bdd100k. "Partially Supervised Training" shows lower-bound performance using partial labels in a centralized setting. "Vanilla Semi-Supervised Federated Learning" and "Our FedSTO" demonstrate improved performance with non-IID federated data. FedSTO approaches the "Fully Supervised Training" upper-bound performance under full label use in a centralized setting. The x-axis shows the number of labeled examples, and the y-axis displays the mean average precision (mAP@0.5) on the test set.
• Figure 3: Performance of various methods on the BDD100K dataset yu2020bdd100k, with the server containing labeled data for the "Cloudy" category and 3 clients having unlabeled data for "Rainy", "Snowy", and "Overcast" categories. Baseline SSFL (red hatched boxes) struggles in comparison to centralized learning (bars in green dotted boxes). "Fully Supervised" and "Partially Supervised" refer to training a centralized model with the complete labeled dataset and only the "Cloudy" labeled data, respectively.
• Figure 4: (a) Performance of various methods on the SODA10m dataset in an IID setting, (b) Average performance across different weather conditions using unlabeled data from the SODA10m dataset.
• Figure : FedSTO Algorithm within the SSFOD Framework