在grid_anchor_generator.py中,此代码生成的锚点与更快的RCNN纸张实现不同。
在Google Object Detection API中,在要素地图位置[0, 0]
中计算的锚点为:
[[ -83.50967407 -48.25482941 97.50967407 42.25482941]
[-174.01934814 -93.50965881 188.01934814 87.50965881]
[-355.03869629 -184.01931763 369.03869629 178.01931763]
[ -57.00000763 -67. 71.00000763 60.99999619]
[-121.00001526 -131. 135.00001526 124.99999237]
[-249.00003052 -259. 263.00003052 252.99998474]
[ -38.25483704 -93.50965881 52.25483704 87.50965881]
[ -83.50967407 -184.01931763 97.50967407 178.01931763]
[-174.01933289 -365.03866577 188.01933289 359.03866577]]
在更快的rcnn论文的作者实现中,锚点是:
[[ -83., -39., 100., 56.],
[-175., -87., 192., 104.],
[-359., -183., 376., 200.],
[ -55., -55., 72., 72.],
[-119., -119., 136., 136.],
[-247., -247., 264., 264.],
[ -35., -79., 52., 96.],
[ -79., -167., 96., 184.],
[-167., -343., 184., 360.]]
更快的rcnn论文中的scale和aspect_ratios是:
scale=[8, 16, 32]
aspect_ratios=[0.5, 1.0, 2.0]
Google Object Detection API中的scale和aspect_ratios是:
scale=[0.5, 1.0, 2.0]
aspect_ratio=[0.5, 1.0, 2.0]