Research On Domain Adaptive Object Detection And Search In Open World

Object detection and search tasks have wide application scenarios.However,existin g detection and search methods are designed with the closed environment,which ignore s the application demands of the open world.Following the assumption that the training and testing sets are similar,the traditional object detection model ignores the complexity of the application scenarios and lacks the adaptability to different environments.Conve ntional object search algorithms only consider the foreground information,which ignore s the background in the image and neglects the importance of detection.This dissertatio n focuses on the study of object detection and search in the open world.With domain ad aptive object detection and multi-task integrated learning,the adaptability of different e nvironments is improved.In general,the following contributions are presented in this di ssertation.(1)A source guided multi-adversarial cross domain detectorThe domain disparity of images caused by environmental changes is ignored by tra ditional object detection model,which also influences the feature of the detector.Thus,t he traditional detector lacks environmental adaptability.To this end,the multi-adversari al domain adaptive model is proposed to minimize the domain disparity in different leve ls of the model.In detail,for the feature confusion of convolutional blocks,the hierarchi cal domain alignment module is proposed with a scale reduction module,which improv es the training efficiency.Besides,in order to get the semantic alignment on the instance level,the aggressive feature alignment module is introduced with WGRL to enhance th e training of the hard sample.With the training of multi-adversarial,the domain disparit y is reduced,and the adaptability is improved.The multi-adversarial learning reduces the domain disparity but ignores domain ada ptability.In domain adaptive object detection task,the distribution of the target domain i s mussy due to the lack of label.With adversarial learning,the source distribution is clos e to the target,such that its feature discrimination is influenced,which further hurt the d omain adaptability.To reduce the domain disparity and preserve domain adaptability,th e source-guided training strategy is introduced.In detail,with knowledge distillation,th e source-pretrained model provides the distribution information to supervise the training phase,which helps the model to preserve domain adaptability.Besides,to get a better marginal distribution of features,the dual discriminator is proposed for the divide of for eground and background samples.Finally,the source-guided model can reduce the dom ain disparity while preserving domain adaptability.(2)A partial alignment asymmetric tri-way detection networkAdversarial learning can reduce the domain disparity,but it also influence the doma in adaptability.Traditional cross-domain object detectors with shared parameters can ac cumulate source distortion caused by adversarial learning,which further influences dom ain adaptability.Meanwhile,the domain disparity can not be completely eliminated,the remaining disparity also influences the model.To this end,the asymmetric tri-way netw ork is proposed.In detail,the ancillary net which is independent of the chief net can avo id the influence of source distortion,which can preserve the domain adaptability.The an cillary net can also provide the ancillary target sample,which adjusts the decision bound ary and make the detector adapt to the target domain.Finally,the model can get better p erformance on cross-domain object detection tasks.However,adversarial learning aligns the whole features of the source and target do mains.Because a specific environment has its specific information,aligning the whole f eature may force the target feature to contain source-specific information,which is unre asonable and causes negative transfer.To this end,the partial alignment training strategy is introduced.In detail,in order to locate the domain-invariant feature during the trainin g phase,inter-adversarial learning is proposed.To decouple the domain-invariant and so urce-specific features,the intra-adversarial is introduced with mutual information theory.With the combination of inter-adversarial and intra-adversarial learning,the model deco uples the domain-invariant and source-specific features and prevents negative transfer.(3)A “divide and conquer” end-to-end “detection-match” integrated network f or object searchThe conventional object search model ignores the background information and desi gns its retrieval part independently.Thus,the cooperation of detection and retrieval is n ot so well.To this end,an end-to-end “detection-match” integrated network is proposed,with end-to-end multi-task training,the detection and retrieval can cooperate better.For the sample insufficient problem of the end-to-end model,the integrated network used th e Simease structure,which expands the number of input samples and enriches the pairin g method.In order to train the metric loss,the online pairing loss and hard example prio rity loss are proposed with the feature dictionary.By combining the improvement of the loss function and net structure.The model can be trained end-to-end and get great perfor mance.A further study found that the detection and retrieval tasks have different focuses.T he detection cares about the differences between foreground and background,while the retrieval tasks are concerned about the semantic differences of foreground samples.Ho wever,the end-to-end model mentioned before gets the detection and retrieval features with a single fc layer,which increases the training difficulty of the model.To this end,t he thought of “divide and conquer” is proposed for the new net structure.In detail,by s haring the low-level net and dividing the high-level,the accuracy of both detection and r etrieval is increased.Besides,due to insufficient sample,the hard example priority loss i s hard to be trained,thus,the class center guided hard example priority loss is introduce d by replacing the parameter with the class center to improve the training efficiency.Wit h the improvement of both net structure and loss function,the model achieves better per formance.