dense-equi-torch

Torch7 implementation of "Unsupervised object learning from dense equivariant image labelling"

Note: I am working on training/test regressor code to make cleaner version.
but pretraining the network for latent space mapping works prefectly now. (don't worry :))

Prerequisites

• Torch7
• thinplatspline
• python 2.7
• other torch packages (xlua, display, hdf5, image ...)

loarocks install display
loarocks install hdf5
loarocks install image
loarocks install xlua

Usage

first, download CelebA dataset (here).

<data_path>
           |-- image 1
           |-- image 2
           |-- iamge 3 ...

To train the feature extractor(CNN):

1. change options in "script/opts.lua" and "data/gen_tps.py"
2. do "th pretrain.lua"
>> pretrained model will saved in 'repo/pretrain/'

To train the regressor(mlp):

1. change options in "script/opts.lua" and "data/gen_reg.lua"
2. do "th regtrain.lua"
>> trained regressor will saved in 'repo/regressor/'

To test the regressor(mlp):

1. change options in "script/opts.lua"
2. do "th regtest.lua"
>> test image wih landmarks will be saved in 'repo/test'

Results

(1) mapping on the latent space

• Red : left-mouth
• Purple : right-mouth
• Green : nose
• Blue : left-eye
• Orange : right-eye

(https://plot.ly/~stellastra666/156/)
(https://plot.ly/~stellastra666/162/)

(2) landmark detection on CelebA

1. good case (red: predict / green: GT)

2. badcase

(3) Performance Benchmark

1. Original paper

nLandmark	regressor training	IOD error
10	CelebA	6.32
30	CelebA	5.76
50	CelebA	5.33

2. My code

nLandmark	regressor training	Iter(reg)	MSE	IOD error
100	CelebA	5K	3.15	5.71
100	CelebA	50K	3.31	5.67

(4) Effect of training data when fine-tuning regressor(mlp)

Training images	learning iter	training loss	MSE	IOD error
10	1K	0.04	5.67	9.97
50	1K	0.09	4.73	8.07
100	1K	0.13	4.42	8.13
2000	2K	0.18	3.38	6.28
5000	3K	0.20	3.36	5.84
15000	5K	0.21	3.15	5.71
15000	50K	0.21	3.31	5.67

ACKNOWLEDGEMENT

Thank James for kindly answering my inquries and providing pieces of matlab code :)

Author

MinchulShin / @nashory