Source-Free Object Detection by Learning to Overlook Domain Style

Abstract

Source-free object detection (SFOD) needs to adapt a detector pre-trained on a labeled source domain to a target domain, with only unlabeled training data from the target domain. Existing SFOD methods typically adopt the pseudo labeling paradigm with model adaption alternating between predicting pseudo labels and fine-tuning the model. This approach suffers from both unsatisfactory accuracy of pseudo labels due to the presence of domain shift and limited use of target domain training data. In this work, we present a novel Learning to Overlook Domain Style (LODS) method with such limitations solved in a principled manner. Our idea is to reduce the domain shift effect by enforcing the model to overlook the target domain style, such that model adaptation is simplified and becomes easier to carry on. To that end, we enhance the style of each target domain image and leverage the style degree difference between the original image and the enhanced image as a self-supervised signal for model adaptation. By treating the enhanced image as an auxiliary view, we exploit a studentteacher architecture for learning to overlook the style degree difference against the original image, also characterized with a novel style enhancement algorithm and graph alignment constraint. Extensive experiments demonstrate that our LODS yields new state-of-the-art performance on four benchmarks.

SGDA: Towards 3-D Universal Pulmonary Nodule Detection via Slice Grouped Domain Attention

“…To the best of our knowledge, this is the first work that deals with SFUDA in the pulmonary nodule detection task. Recent SFUDA object detection works utilize the pseudo labeling strategy [18] associated with feature alignment or sample generation [19], [20], [21], which are consistent with the works in SFUDA image classification and segmentation.…”

“…Nonetheless, its ability to distinguish the nodules from other tissues is weaker. It should be noted that our re-implementation of the LODS [20] removes the style enhancement module. This is because the reconstruction of the CT images itself is already a challenging problem.…”

“…Thus the discriminative features of the nodules are not fully explored and the classification of the nodules is not satisfactory. Moreover, after adding the graph based image-level alignment, the performance of the LODS [20] even decreases. We believe the reason is that, in small-sized pulmonary nodule detection, the image-level features involves much more voxel information than the instancelevel features and its alignment causes disturbance to the instance-level alignment.…”

“…The HCL [19] utilizes historical contrastive learning to exploit the source domain and learn the target representations. LODS [20] treats the difference between the original target image and the style enhanced one as an auxiliary supervision for model adaptation. In [21], the authors introduce a more realistic online SFUDA setting, and design a cross-attention transformerbased memory module for the target representation learning.…”

AdaptGuard: Defending Against Universal Attacks for Model Adaptation

“…This new paradigm [55,28,20,2,50,9,42] quickly becomes popular among the transfer learning field thanks to the appealing privacy-preserving property and the competitive performance to source-data-dependent methods. Nowadays, the model adaptation scheme has been widely applied to various tasks, e.g., image classification [25,55,28,17], semantic segmentation [29,19,57], and object detection [54,22].…”