2019/05/14 by Ivan Krešo, Krešo, Ivan, Josip Krapac +3 · 1 citation
Computer Science · #Advanced Image and Video Retrieval Techniques #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences
paper · pdf · doi:10.48550/arxiv.1905.05661
openalex publication_date 2019/05/14 · openalex created_date 2022/07/29 · openalex updated_date 2026/07/28
Recent progress of deep image classification models has provided great\npotential to improve state-of-the-art performance in related computer vision\ntasks. However, the transition to semantic segmentation is hampered by strict\nmemory limitations of contemporary GPUs. The extent of feature map caching\nrequired by convolutional backprop poses significant challenges even for\nmoderately sized Pascal images, while requiring careful architectural\nconsiderations when the source resolution is in the megapixel range. To address\nthese concerns, we propose a novel DenseNet-based ladder-style architecture\nwhich features high modelling power and a very lean upsampling datapath. We\nalso propose to substantially reduce the extent of feature map caching by\nexploiting inherent spatial efficiency of the DenseNet feature extractor. The\nresulting models deliver high performance with fewer parameters than\ncompetitive approaches, and allow training at megapixel resolution on commodity\nhardware. The presented experimental results outperform the state-of-the-art in\nterms of prediction accuracy and execution speed on Cityscapes, Pascal VOC\n2012, CamVid and ROB 2018 datasets. Source code will be released upon\npublication.\n