2019/09/21 by Kemal Öksüz, Oksuz, Kemal, Barış Can Çam +5
Computer Science · Medicine · #Advanced Neural Network Applications #COVID-19 diagnosis using AI #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences
paper · pdf · doi:10.48550/arxiv.1909.09777
openalex publication_date 2019/09/21 · openalex created_date 2022/09/27 · openalex updated_date 2026/07/28
Two-stage deep object detectors generate a set of regions-of-interest (RoI)\nin the first stage, then, in the second stage, identify objects among the\nproposed RoIs that sufficiently overlap with a ground truth (GT) box. The\nsecond stage is known to suffer from a bias towards RoIs that have low\nintersection-over-union (IoU) with the associated GT boxes. To address this\nissue, we first propose a sampling method to generate bounding boxes (BB) that\noverlap with a given reference box more than a given IoU threshold. Then, we\nuse this BB generation method to develop a positive RoI (pRoI) generator that\nproduces RoIs following any desired spatial or IoU distribution, for the\nsecond-stage. We show that our pRoI generator is able to simulate other\nsampling methods for positive examples such as hard example mining and prime\nsampling. Using our generator as an analysis tool, we show that (i) IoU\nimbalance has an adverse effect on performance, (ii) hard positive example\nmining improves the performance only for certain input IoU distributions, and\n(iii) the imbalance among the foreground classes has an adverse effect on\nperformance and that it can be alleviated at the batch level. Finally, we train\nFaster R-CNN using our pRoI generator and, compared to conventional training,\nobtain better or on-par performance for low IoUs and significant improvements\nwhen trained for higher IoUs for Pascal VOC and MS COCO datasets. The code is\navailable at: https://github.com/kemaloksuz/BoundingBoxGenerator.\n