Exploiting Low-confidence Pseudo-labels for Source-free Object Detection

Zhihong Chen,Zilei Wang,Yixin Zhang

DOI: https://doi.org/10.1145/3581783.3612273

2023-10-19

Abstract:Source-free object detection (SFOD) aims to adapt a source-trained detector to an unlabeled target domain without access to the labeled source data. Current SFOD methods utilize a threshold-based pseudo-label approach in the adaptation phase, which is typically limited to high-confidence pseudo-labels and results in a loss of information. To address this issue, we propose a new approach to take full advantage of pseudo-labels by introducing high and low confidence thresholds. Specifically, the pseudo-labels with confidence scores above the high threshold are used conventionally, while those between the low and high thresholds are exploited using the Low-confidence Pseudo-labels Utilization (LPU) module. The LPU module consists of Proposal Soft Training (PST) and Local Spatial Contrastive Learning (LSCL). PST generates soft labels of proposals for soft training, which can mitigate the label mismatch problem. LSCL exploits the local spatial relationship of proposals to improve the model's ability to differentiate between spatially adjacent proposals, thereby optimizing representational features further. Combining the two components overcomes the challenges faced by traditional methods in utilizing low-confidence pseudo-labels. Extensive experiments on five cross-domain object detection benchmarks demonstrate that our proposed method outperforms the previous SFOD methods, achieving state-of-the-art performance.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper aims to address the Source-Free Object Detection (SFOD) problem in Unsupervised Domain Adaptation Object Detection (UDAOD). Specifically, existing SFOD methods typically use threshold-based pseudo-labeling methods during the adaptation phase, which are often limited to high-confidence pseudo-labels, leading to information loss. To address this issue, the authors propose a new method that fully utilizes pseudo-labels by introducing two thresholds: high confidence and low confidence.

Specifically:
1. **High-Confidence Pseudo-Labels**: For pseudo-labels with confidence higher than the high threshold, traditional methods are used to directly train with these pseudo-labels.
2. **Low-Confidence Pseudo-Labels**: For pseudo-labels with confidence between the high and low thresholds, a module named Low-Confidence Pseudo-Label Utilization (LPU) is introduced for training. The LPU module consists of two sub-modules:
- **Proposal Soft Training (PST)**: Generates soft labels for proposal training to alleviate the label mismatch problem caused by traditional IoU-based label assignment methods.
- **Local Spatial Contrastive Learning (LSCL)**: Utilizes local spatial relationships to optimize proposal feature representations, enhancing the model's understanding of relationships between neighboring proposals.

By combining the PST and LSCL components, this method overcomes the challenges faced by traditional methods in utilizing low-confidence pseudo-labels. Experimental results show that this method achieves better performance than existing SFOD methods on 5 cross-domain object detection benchmark datasets.