Domain Adaptive Object Detection For Different Weather Scenarios

In recent years,object detection has made significant research progress and has been widely used in daily life,including automatic driving,intelligent traffic,and so on.However,when the training data and test data come from different lighting or weather conditions,the performance of object detection is often poor because of domain shift.To improve the generalization ability of the object detector for different weather scenarios,domain adaptive object detection(DAOD)has received a lot of attention.For DAOD,recent advances mainly explore aligning feature-level distributions between the source and single-target domain,which may neglect the impact of domain-specific information existing in the aligned features.In many data-critical scenarios with privacy issues,training images could not be transferred from source domain to target domain.And the data in the target domain is not a single weather scenario and may contain compound scenarios,which leads to a more challenging task of DAOD.Towards the domain-specific information in the feature and compound weather scenarios,we propose a novel disentangled method based on vector decomposition.Firstly,domain-invariant representations are separated from the input by adversarial training.Secondly,domain-specific representations are introduced as the differences between the input and domain-invariant representations.And we can disentangle domaininvariant representations from domain-specific representations.And orthogonality and completeness of disentanglement are guaranteed.It solves the DAOD based on domain invariant features in the single-and compound-target case.Experimental results show our method obtains a significant performance gain over baseline methods.On the basis of the above research,we propose a Pseudo-Label method based on Score Net towards the source data-free DAOD.We use the Score Net to choose more positive categories.It corrects the confidence of the bounding boxes by the image-level classification.The quality of pseudo-labels is improved and we fine-tune the source model on the target domain.It solves the source data-free DAOD under privacy restrictions.We evaluate the method on different weather scenarios.Experimental results demonstrate the effectiveness of Score Net and the state-of-the-art performance.