This research paper presents a novel approach to source-free object detection (SFOD), which involves using a pre-trained model on a source domain and unlabeled target domain data. The authors propose the use of low-confidence pseudo-labels in addition to high-confidence ones to avoid loss of information. The introduced method includes a Low-confidence Pseudo-labels Utilization (LPU) module, which consists of Proposal Soft Training (PST) and Local Spatial Contrastive Learning (LSCL). PST mitigates the label mismatch problem, while LSCL improves the model’s ability to differentiate between spatially adjacent proposals. The method demonstrated superior performance over traditional SFOD methods in tests on five cross-domain object detection benchmarks.
Publication date: 20 Oct 2023
Project Page: https://doi.org/10.1145/3581783.3612273
Paper: https://arxiv.org/pdf/2310.12705
