Action-Recognition

从零开始的Action-Recognition —— 基于mxnet gluon 和 gluoncv 的行为识别

./model_zoo 模型文件
- model_store.py 卡太小又下不完k400 只好蹭gluoncv的预训练数据这样子
- model_zoo.py 模型库的头文件网络结构就是要整整齐齐
- init.py import用写了新模型记得填一下
- ~~simple.py 简单分类为了项目框架写的模型文件示例 demo [drop]~~
- ~~test_1E4.py 有了前背景想法以后的粗糙实现已经[drop]~~
- ~~MSEloss_vgg.py 验证学习一致性有没有提高 [drop]~~
- ~~inception_v3_k400ft.py 使用gluoncv inceptionv3的参数作为用分类做视频的baseline [drop]~~
- ~~actionrec_inceptionv3.py 使用gluoncv inceptionv3的参数分类 [drop]~~
- ~~inceptionv3_LSTM.py [drop]~~
- resnet18_v1b_ucf101.py 使用了gluoncv实现的reset 作为分类baseline、复现双流、TSN、前背景的backbone [2D-ResNet/2-stream/TSN/bfgs]
- C3D.py 调了一下 [C3D]
- Res3D.py 调了一下 [3D-ResNet]
- R21D.py 调了一下 [R(2+1)D-34]
- F_stCN.py 简单调了一下一篇比较早的[FstCN]
- ~~P3D.py 调P3D 没搞好 [drop]~~
- mx_c3d.py gluoncv实现的c3d [C3D]
- mx_c3d_base.py 基于 mx_c3d.py改的本来是打算用来和其它无监督算法基于C3D对比的 [C3D&decoder]
- r2plus1d.py gluoncv实现的 [R(2+1)D]
- r2plus1d_base.py 基于 r2plus1d.py改的用来作为无监督算法的backbone [R(2+1)D&decoder]
- ECO.py 调了一下[ECO] 太依赖预训练了
- mx_i3d_resnet.py 用了一下gluoncv的[I3D]
./ucf101_bgs 数据集加载

主要是twostream.py 基于原本gluoncv.data.ucf101修改的双流版本的UCF101和HMDB51的数据集加载脚本
pre_ucf101.py 预处理数据集用
BgsDecomposer.py 放在bgslibrary目录下运行用于从视频中分割出前景的脚本
test_model.py 测试单个模型用的脚本
test_twostream_skip.py 测试双流模型用的脚本
train_%d_%d%s.py 实验训练脚本
~~其它奇怪的脚本无视~~

实验

train_1_%d 基于双流结构的实验包括分类、TSN、前背景、I3D
train_2_%d 训练C3D的配置
train_3_%d 训练ECO
train_4_1&train_4_2 训练各种3D卷积、R(2+1)D
train_4_3& train_4_4 R(2+1)D作base+反卷积实现3D自编码
train_5_%d R(2+1)D作base+反卷积实现重构和预测
train_6_%d 尝试了一下变换自编码器AET 没收敛
train_7_%d R(2+1)D作base+反卷积 +resnet34TSN 基于灰度一致性的对抗训练

Name		Name	Last commit message	Last commit date
Latest commit History 20 Commits
model_zoo		model_zoo
ucf101_bgs		ucf101_bgs
.gitignore		.gitignore
BgsDecomposer.py		BgsDecomposer.py
DualBlock.py		DualBlock.py
DualBlock_fc.py		DualBlock_fc.py
README.md		README.md
loss.py		loss.py
mytrain_recognizer.py		mytrain_recognizer.py
mytrain_recognizer1.py		mytrain_recognizer1.py
pre_ucf101.py		pre_ucf101.py
test_1_1.py		test_1_1.py
test_1_2.py		test_1_2.py
test_fgsbgs.py		test_fgsbgs.py
test_lstm.py		test_lstm.py
test_model.py		test_model.py
test_model_inception.py		test_model_inception.py
test_model_inception_diff.py		test_model_inception_diff.py
test_mseloss.py		test_mseloss.py
test_recognizer.py		test_recognizer.py
test_twostream.py		test_twostream.py
test_twostream_skip.py		test_twostream_skip.py
test_twostream_skip_i3d.py		test_twostream_skip_i3d.py
train_1_1_renet18v2.py		train_1_1_renet18v2.py
train_1_2.py		train_1_2.py
train_1_3.py		train_1_3.py
train_1_4.py		train_1_4.py
train_1_5_newlength.py		train_1_5_newlength.py
train_1_6.py		train_1_6.py
train_1_7_newlength.py		train_1_7_newlength.py
train_1_8_i3d.py		train_1_8_i3d.py
train_2_1.py		train_2_1.py
train_2_2.py		train_2_2.py
train_2_3.py		train_2_3.py
train_3_1.py		train_3_1.py
train_3_2.py		train_3_2.py
train_3_3.py		train_3_3.py
train_4_1.py		train_4_1.py
train_4_2.py		train_4_2.py
train_4_3.py		train_4_3.py
train_4_4.py		train_4_4.py
train_4_5.py		train_4_5.py
train_5_1.py		train_5_1.py
train_5_2.py		train_5_2.py
train_5_3.py		train_5_3.py
train_5_4.py		train_5_4.py
train_5_5.py		train_5_5.py
train_5_6.py		train_5_6.py
train_5_6_1.py		train_5_6_1.py
train_5_7.py		train_5_7.py
train_5_8.py		train_5_8.py
train_5_9.py		train_5_9.py
train_5_9_1.py		train_5_9_1.py
train_6_1.py		train_6_1.py
train_6_1old.py		train_6_1old.py
train_6_2.py		train_6_2.py
train_7_1.py		train_7_1.py
train_7_2.py		train_7_2.py
train_7_3.py		train_7_3.py
train_7_4.py		train_7_4.py
train_7_5.py		train_7_5.py
train_8_02.py		train_8_02.py
train_8_1.py		train_8_1.py
train_8_2.py		train_8_2.py
train_fgsbgs.py		train_fgsbgs.py
train_model.py		train_model.py
train_recognizer.py		train_recognizer.py
train_recognizer_old.py		train_recognizer_old.py
ucf101.py		ucf101.py

bdus/Action-Recognition

Folders and files

Latest commit

History

Repository files navigation

Action-Recognition

实验