2b3bf3eee2475e03885a110e9acaab61-Supplemental.pdf

One20 major reason isthat the filter bank isredundant and contains enough representation power. When21 sampled with different random seeds, the estimator is capable of generating abundant kernels.22 However,wealsoobservetheoscillations ofaccuracy(about 0.5% Top-1accuracyvariance in1023 runs) when the small portion of DoG filters are removed. We perform object detection experiments on COCO 2017 [8] dataset, which41 contains 118K images for training, 5K images for validation, and 20K images for test-dev. We also evaluate our method on semantic segmentation, utilizing the50 widely-used ADE20K [12] dataset. ADE20K covers 150 semantic classes, with 20K images for51 training, 2K images for testing, and 3K for testing.