github:https://github.com/msracver/Flow-Guided-Feature-Aggregation
papers:https://arxiv.org/pdf/1703.10025.pdf
【注意】Flow-Guided-Feature-Aggregation与另一篇文章Deep Feature Flow非常相似,可参考其环境配置。
DFF-github:https://github.com/msracver/Deep-Feature-Flow
MXNet安装-docker_qq_44716044的博客-CSDN博客
#作者声明需要的包
pip install Cython
pip install opencv-python==3.2.0.6
pip install easydict==1.6
python ./fgfa_rfcn/demo.py
【报错】
mxnet.base.NotImplementedForSymbol: Function iadd (namely operator "+=")
找到发生报错的代码片段,将
aggregated_conv_feat += tiled_weight * warp_list[i]
改为下述
aggregated_conv_feat = aggregated_conv_feat + tiled_weight * warp_list[i]
可参考MEGA中作者的CUSTOMIZE.md将自己的数据集制作成标准格式。【待补充】
(1)修改experiments/fgfa_rfcn/cfgs/resnet_v1_101_flownet_imagenet_vid_rfcn_end2end_ohem.yaml中类别数和训练数据集名称
dataset:
NUM_CLASSES: 2#注意应该是自己的类别数+背景
image_set: VID_train_15frames
(2)修改lib/dataset/imagenet_vid.py中类别名称,删去多余类别
self.classes=['__background__',#alwaysindex0
'car']
self.classes_map=['__background__',#alwaysindex0
'n02958343']
python experiments/fgfa_rfcn/fgfa_rfcn_end2end_train_test.py --cfg experiments/fgfa_rfcn/cfgs/resnet_v1_101_flownet_imagenet_vid_rfcn_end2end_ohem.yaml
(1)错误1
mxnet.base.MXNetError: [15:13:33]fail to open fgfa_rfcn_vid-0002.params
报错信息太长只截取了一段,需要修改experiments/fgfa_rfcn/fgfa_rfcn_end2end_train_test中的main函数,去掉第一行注释符号
if __name__ == "__main__":
train_end2end.main()
test.main()
(2)错误2
TypeError: init_params() got an unexpected keyword argument ‘allow_extra’
解决:进入mnxet/module/找到base_module.py文件中 allow_extra=allow_extra这一行删除掉。
(3)错误3
TypeError: _update_params_on_kvstore() takes exactly 4 arguments (3 given)
解决办法:进入对应的module.py文件,相应修改(在服务器中vim显示不全,可通过下载该文件,修改后再上传,上传前需删除原文件)
_update_params_on_kvstore(self._exec_group.param_arrays,
self._exec_group.grad_arrays,
self._kvstore,
self._param_names)
(1)Deep-Feature-Flow文章及代码训练解析_HeyMrYu的博客-CSDN博客_deep feature flow
(2)光流法进行视频目标检测,Flow-Guided-Feature-Aggregation和Deep Feature Flow的论文复现_zyksumm的博客-CSDN博客
(3)Flow-Guided-Feature-Aggregation-的安装配置,demo运行,以及采用少量ILSVRC2015 VID数据集在其中训练_一个想赢的人的博客-CSDN博客
(4)ILSVRC2015_VID数据集详解_lighten-1996的博客-CSDN博客_ilsvrc2015