Meta R-CNN : Towards General Solver for Instance-level Few-shot Learning VOC2012数据集复现结果

Meta R-CNN : Towards General Solver for Instance-level Few-shot Learning VOC2012数据集复现结果

在复现该算法时，由于是使用自定义的数据集和训练规则，遇到了很多踩坑的东西。在此列出我的实验笔记和基础复现结果，给自己做记录也给同样复现的同学们参考。

环境

Requirements
python 3.5
PyTorch = 0.3.1
Torchvision = 0.2.0
cython
pyyaml
easydict
opencv-python
matplotlib
numpy
scipy
tensorboardX
CUDA 8.0 gcc >= 4.9

训练概览

训练1，2阶段 1阶段在base class 上训练 2阶段在base& novel class 上训练
测试在novel class 上测试
lr_decay_step=4 （step to do learning rate decay, unit is epoch）
earning rate decay ratio=0.1
optimizer=sgd

VOC2012标准数据集复现结果

目标检测测试结果 AP50 (overthresh=0.5)
AP for aeroplane = 0.695
AP for bicycle = 0.763
AP for boat = 0.576
AP for bottle = 0.586
AP for car = 0.772
AP for cat = 0.844
AP for chair = 0.450
AP for diningtable = 0.654
AP for dog = 0.851
AP for horse = 0.816
AP for person = 0.762
AP for pottedplant = 0.439
AP for sheep = 0.680
AP for train = 0.698
AP for tvmonitor = 0.719
AP for bird = 0.542 NOVEL
AP for bus = 0.597 NOVEL
AP for cow = 0.661 NOVEL
AP for motorbike = 0.661 NOVEL
AP for sofa = 0.425 NOVEL
Mean AP = 0.6595