2021/08/13 by Omiros Pantazis, Gabriel Brostow, Pantazis, Omiros +5
Computer Science · Biochemistry, Genetics and Molecular Biology · Engineering · #Domain Adaptation and Few-Shot Learning #Genomics and Phylogenetic Studies #Remote-Sensing Image Classification
paper · pdf · doi:10.48550/arxiv.2108.06435
We address the problem of learning self-supervised representations from\nunlabeled image collections. Unlike existing approaches that attempt to learn\nuseful features by maximizing similarity between augmented versions of each\ninput image or by speculatively picking negative samples, we instead also make\nuse of the natural variation that occurs in image collections that are captured\nusing static monitoring cameras. To achieve this, we exploit readily available\ncontext data that encodes information such as the spatial and temporal\nrelationships between the input images. We are able to learn representations\nthat are surprisingly effective for downstream supervised classification, by\nfirst identifying high probability positive pairs at training time, i.e. those\nimages that are likely to depict the same visual concept. For the critical task\nof global biodiversity monitoring, this results in image features that can be\nadapted to challenging visual species classification tasks with limited human\nsupervision. We present results on four different camera trap image\ncollections, across three different families of self-supervised learning\nmethods, and show that careful image selection at training time results in\nsuperior performance compared to existing baselines such as conventional\nself-supervised training and transfer learning.\n