Lately, remarkable advancements of artificial intelligence have been attributed to the integration of self-supervised learning scheme. Despite impressive achievements within NLP, yet SSL in computer vision has not been able to stay on track comparatively. Recently, integration of contrastive learning on top of existing SSL models has established considerable progress in computer vision through which visual SSL models have outperformed their supervised counterparts. Nevertheless, most of these improvements were limited to classification tasks, and also, few works have been dedicated to evaluation of SSL models in real-world scenarios of computer vision, while the majority of works are centered around datasets containing class-wise portrait images, most notably, ImageNet. Consequently, in this work, we have considered dense prediction task of semantic segmentation in security inspection x-ray images to evaluate our proposed model Segmentation Localization. Based upon the model Instance Localization, our model SegLoc has managed to address one of the most challenging downsides of contrastive learning, i.e., false negative pairs of query embeddings. In order to do so, in contrast to baseline model InsLoc, our pretraining dataset is synthesized by cropping, transforming, then pasting already labeled segments from an available labeled dataset, foregrounds, onto instances of an unlabeled dataset, backgrounds. In our case, PIDray and SIXray datasets are considered as labeled and unlabeled datasets, respectively. Moreover, we fully harness labels by avoiding false negative pairs through implementing the idea, one queue per class, in MoCo-v2 whereby negative pairs corresponding to each query are extracted from its corresponding queue within the memory bank. Our approach has outperformed random initialization by 3% to 6%, while having underperformed supervised initialization.
翻译:暂无翻译