目标检测RON网络VOC2007训练测试

源码github

1.RON320 VOC07_test

下载VOC2007数据集，其解压放到RON/data目录下，修改文件夹名VOCdevkit为VOCdevkit2007（与代码中一致）

前提已经下载好模型RON320_VOC0712_VOC07.caffemodel（参照）

在VOCdevkit2007中创建系列文件夹 ./results/VOC2007/Main，否则运行test会报错找不到VOCdevkit2007/results/VOC2007/Main…文件

在RON目录下运行脚本./test_voc07.sh

得到测试结果如下，MAP为0.742

Saving cached annotations to data/VOCdevkit2007/annotations_cache/annots.pkl
AP for aeroplane = 0.7638
AP for bicycle = 0.7953
AP for bird = 0.7472
AP for boat = 0.6666
AP for bottle = 0.5295
AP for bus = 0.8348
AP for car = 0.8339
AP for cat = 0.8582
AP for chair = 0.5569
AP for cow = 0.7916
AP for diningtable = 0.6942
AP for dog = 0.8427
AP for horse = 0.8134
AP for motorbike = 0.8322
AP for person = 0.7609
AP for pottedplant = 0.4906
AP for sheep = 0.7362
AP for sofa = 0.7551
AP for train = 0.8030
AP for tvmonitor = 0.7241
Mean AP = 0.7415

2.RON320 VOC07_train

我是在一路titan显卡上训练

下载预训练模型VGG_ILSVRC_16_layers_fc_reduced.caffemodel，放到data/ImageNet_models，修改脚本train_voc.sh并运行

python tools/train_net.py --gpu 0 \
--solver models/pascalvoc/VGG16/solver.prototxt \
#  只训练VOC2007,默认的--imdb参数就是
--weights data/ImageNet_models/VGG16_layers_fully_conv.caffemodel \
--batchsize 16 \#batchsize 20过大，会cuda out of memory
--iters 10000

注：若出现cudnn报错，因为编译没加cudnn，因此我把prototxt中engine: CUDNN全去掉

注：若output文件夹没有权限，先给其权限

sudo chmod -R 777 ./output

因为时间问题，这儿我是仅训练voc2007，迭代10000次，batchsize为16，训练了7小时，训练结果得到 loss=1.195

测试map，运行命令

python ./tools/test_net.py --gpu 0 \
--def models/pascalvoc/VGG16/test320cudnn.prototxt \
--net output/default/voc_2007_trainval/RON_320_iter_10000.caffemodel \
--imdb voc_2007_test

结果为 map=0.61,达到作者效果的82%。时间充足可以尝试按照官方给的迭代120000次，batchsize为20，VOC2007+2012训练

AP for aeroplane = 0.6431
AP for bicycle = 0.7331
AP for bird = 0.5981
AP for boat = 0.4798
AP for bottle = 0.3230
AP for bus = 0.6838
AP for car = 0.7294
AP for cat = 0.7936
AP for chair = 0.3810
AP for cow = 0.6029
AP for diningtable = 0.4608
AP for dog = 0.6924
AP for horse = 0.7696
AP for motorbike = 0.6863
AP for person = 0.6396
AP for pottedplant = 0.3565
AP for sheep = 0.5833
AP for sofa = 0.6703
AP for train = 0.7615
AP for tvmonitor = 0.6178
Mean AP = 0.6103