Rethinking confidence scores for source-free unsupervised domain adaptation

Source-free unsupervised domain adaptation (SFUDA) aims to achieve target domain predictions through a source model instead of source data. One of the representative ideas for the SFUDA problem is to apply a self-supervised pseudo-labeling strategy (SSPL) to achieve target domain adaptation, but it is prone to being plagued by noisy labels, which can lead to negative transfer. Therefore, many methods attempt to improve the SSPL and leverage confidence scores to weaken the impact of low-confidence samples on the model, which are potentially noisy samples. However, they are unable to completely overcome the problem of noisy labels because the pseudo-labels of high-confidence samples may also be incorrect. Besides, they rarely allow low-confidence samples to be added to training, which can lead to sample selection bias and thus limit the model generalization ability. In this work, we propose information re-exploitation based on confidence scores (RECS) for the SFUDA problem, in which we rethink the information brought by confidence scores and take advantage of them to solve the shortcomings of the improved SSPL. Specifically, we realize cross-domain target adaptation by the symmetric SSPL with dual denoising, and reduce the intra-domain distribution discrepancy by the discriminative class-balanced feature alignment. In this way, the model robustness and generalization are enhanced. Extensive experiments conducted on three standard datasets have demonstrated the effectiveness and superiority of our proposed method. The code is available at https://github.com/lingyuxuan1234/RECS.