Abstract:Effective object detection in autonomous vehicles is challenged by deployment in diverse and unfamiliar environments. Online Source-Free Domain Adaptation (O-SFDA) offers model adaptation using a stream of unlabeled data from a target domain in an online manner. However, not all captured frames contain information beneficial for adaptation, especially in the presence of redundant data and class imbalance issues. This paper introduces a novel approach to enhance O-SFDA for adaptive object detection through unsupervised data acquisition. Our methodology prioritizes the most informative unlabeled frames for inclusion in the online training process. Empirical evaluation on a real-world dataset reveals that our method outperforms existing state-of-the-art O-SFDA techniques, demonstrating the viability of unsupervised data acquisition for improving the adaptive object detector.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to achieve effective object detection in autonomous vehicles. Specifically, when these vehicles are deployed in variable and unfamiliar environments, the performance of pre - trained object detection models will decline significantly. The paper focuses on Online Source - Free Domain Adaptation (O - SFDA), which is a method of model adaptation using unlabeled data streams in the target domain without source data. However, existing O - SFDA methods have some limitations, such as: 1. **Data Redundancy**: Not all captured frames contain information that is beneficial for model adaptation, especially when there are problems of redundant data and class imbalance. 2. **High Computational Cost**: Using every frame for adaptation will lead to high computational costs and may exacerbate the class imbalance problem, making the detection frequency of common objects (such as cars and people) higher and that of rare objects (such as motorcycles) lower, thus affecting the overall detection performance. To solve these problems, the paper proposes a new method based on unsupervised data acquisition, which dynamically identifies and integrates the most informative unlabeled frames and frames containing rare classes through incremental online clustering. This method aims to reduce the use of computational resources while enhancing the adaptability of the object detection model in unknown deployment environments. Specifically, the paper proposes a two - stage data acquisition strategy: 1. **Acquisition Unsim Frame (AUF)**: By comparing the similarity between new frames and existing cluster centers, frames that are significantly different from the existing data are selected as key frames. 2. **Acquisition Rare Category (ARC)**: Through an additional clustering mechanism, frames containing rare classes are given a second chance to be selected to solve the class imbalance problem. Through these two - stage data acquisition strategies, the method in the paper can improve the adaptability and performance of the object detection model while reducing computational resources. Experimental results show that this method outperforms the existing state - of - the - art O - SFDA techniques on multiple datasets.

Improving Online Source-free Domain Adaptation for Object Detection by Unsupervised Data Acquisition

A Free Lunch for Unsupervised Domain Adaptive Object Detection without Source Data

Unsupervised Domain Adaptation Approach for Vision-Based Semantic Understanding of Bridge Inspection Scenes Without Manual Annotations

Source-free Domain Adaptive Object Detection in Remote Sensing Images

Source-free Domain Adaptation for Video Object Detection Under Adverse Image Conditions

Source-Free Domain Adaptation for Semantic Segmentation

SF-FSDA: Source-Free Few-Shot Domain Adaptive Object Detection with Efficient Labeled Data Factory

Multi-source-free Domain Adaptive Object Detection

Simplifying Source-Free Domain Adaptation for Object Detection: Effective Self-Training Strategies and Performance Insights

Proposal-Level Unsupervised Domain Adaptation for Open World Unbiased Detector

Source-Free Domain Adaptation for YOLO Object Detection

Unsupervised Domain Adaptation for Object Detection via Cross-Domain Semi-Supervised Learning

C2FDA: Coarse-to-Fine Domain Adaptation for Traffic Object Detection

Source data‐free domain adaptation of object detector through domain‐specific perturbation

Domain Adaptive Object Detection for Autonomous Driving under Foggy Weather

Multi-level domain perturbation for source-free object detection in remote sensing images

SSDA-YOLO: Semi-supervised domain adaptive YOLO for cross-domain object detection

Self-Guided Adaptation: Progressive Representation Alignment for Domain Adaptive Object Detection

Unsupervised Domain Adaptation for Remote-Sensing Vehicle Detection Using Domain-Specific Channel Recalibration

Uncertainty-Aware Active Domain Adaptive Salient Object Detection

Domain Adaptation based Object Detection for Autonomous Driving in Foggy and Rainy Weather