@@ -4,47 +4,48 @@ NOTE: This is not official implementation. Original paper is [DeepPose: Human Po
44
55# Requirements
66
7- - Python 2.7.11+
8-
9- - [ Chainer 1.5+] ( https://github.com/pfnet/chainer ) (Neural network framework)
7+ - Python 3.5.1+
8+ - [ Chainer 1.13.0+] ( https://github.com/pfnet/chainer )
109 - numpy 1.9+
11- - scipy 0.16+
12- - scikit-learn 0.15+
13- - OpenCV 2.4+
10+ - scikit-image 0.11.3+
11+ - OpenCV 3.1.0+
12+
13+ I strongly recommend to use Anaconda environment. This repo may be able to be used in Python 2.7 environment, but I haven't tested.
1414
1515## Installation of dependencies
1616
1717```
1818pip install chainer
1919pip install numpy
20- pip install scipy
21- pip install scikit-learn
20+ pip install scikit-image
2221# for python3
2322conda install -c https://conda.binstar.org/menpo opencv3
2423# for python2
2524conda install opencv
2625```
2726
28- # Data preparation
27+ # Dataset preparation
2928
3029```
31- bash shells /download.sh
32- python scripts /flic_dataset.py
33- python scripts /lsp_dataset.py
34- python scripts /mpii_dataset.py
30+ bash datasets /download.sh
31+ python datasets /flic_dataset.py
32+ python datasets /lsp_dataset.py
33+ python datasets /mpii_dataset.py
3534```
3635
37- This script downloads FLIC-full dataset (< http://vision.grasp.upenn.edu/cgi-bin/index.php?n=VideoLearning.FLIC > ) and perform cropping regions of human and save poses as numpy files into FLIC-full directory. Same processes are performed for LSP, MPII datasets.
36+ - [ FLIC-full dataset] ( http://vision.grasp.upenn.edu/cgi-bin/index.php?n=VideoLearning.FLIC )
37+ - [ LSP Extended dataset] ( http://www.comp.leeds.ac.uk/mat4saj/lspet_dataset.zip )
38+ - ** MPII dataset**
39+ - [ Annotation] ( http://datasets.d2.mpi-inf.mpg.de/leonid14cvpr/mpii_human_pose_v1_u12_1.tar.gz )
40+ - [ Images] ( http://datasets.d2.mpi-inf.mpg.de/andriluka14cvpr/mpii_human_pose_v1.tar.gz )
3841
3942## MPII Dataset
4043
4144- [ MPII Human Pose Dataset] ( http://human-pose.mpi-inf.mpg.de/#download )
4245- training images: 18079, test images: 6908
43-
4446 - test images don't have any annotations
4547 - so we split trining imges into training/test joint set
4648 - each joint set has
47-
4849- training joint set: 17928, test joint set: 1991
4950
5051# Start training
@@ -72,49 +73,12 @@ bash shells/train_mpii.sh
7273### GPU memory requirement
7374
7475- AlexNet
75-
7676 - batchsize: 128 -> about 2870 MiB
7777 - batchsize: 64 -> about 1890 MiB
7878 - batchsize: 32 (default) -> 1374 MiB
79-
8079- ResNet50
81-
8280 - batchsize: 32 -> 6877 MiB
8381
84- # Visualize Filters of 1st conv layer
85-
86- - Go to result dir of a model
87- - ` python ../../scripts/draw_filters.py `
88-
89- # Visualize Prediction
90-
91- ## Example
92-
93- ### Prediction and visualize them and calc mean errors
82+ # Prediction
9483
95- ```
96- python scripts/evaluate_flic.py \
97- --model results/AlexNet_2015/AlexNet.py \
98- --param results/AlexNet_2015/AlexNet_epoch_400.chainermodel \
99- --datadir data/FLIC-full
100- --gpu 0 \
101- --batchsize 128 \
102- --mode test
103- ```
104-
105- ### Tile some randomly selected result images
106-
107- ```
108- python scripts/evaluate_flic.py \
109- --model results/AlexNet_2015/AlexNet_flic.py \
110- --param results/AlexNet_2015/AlexNet_epoch_450.chainermodel \
111- --mode tile \
112- --n_imgs 25
113- ```
114-
115- ### Create animated GIF to intuitively compare predictions and labels
116-
117- ```
118- cd results/AlexNet_2015
119- bash ../../scripts/create_anime.sh test_450_tiled_pred.jpg test_450_tiled_label.jpg test_450.gif
120- ```
84+ Will add some tools soon
0 commit comments