bknyaz/beauty_vision

Recognition of human faces attractiveness (SCUT-FBP dataset) in Python

beauty_vision

Recognition of human faces attractiveness (the SCUT-FBP dataset [1])

Overview

The SCUT-FBP dataset [1] contains 500 samples (images), for each image there is rating in the range (1,5) measuring beauty of an Asian female face.

Support Vector Regression (SVR) is trained on top of different features (in some cases projected by PCA with 50 components). Average Pearson correlation (PC) for 5 independent 10-fold cross validation tests is reported as in [1].

In all experiments images are first resized to (224,294), then central crop (224,224) is taken.

Results

Model	Code	Avg PC for 5 tests
Combined features + PCA + SVR [1]	-	0.6433
ConvNet [1]	-	0.8187
16 random filters + PCA50+ rbf SVR	[beauty_baseline_random] (beauty_baseline_random.py)	0.642
16 random filters + linear SVR	[beauty_baseline_random] (beauty_baseline_random.py)	0.646
24 random filters + linear SVR	[beauty_baseline_random] (beauty_baseline_random.py)	0.660
24 Gabor filters + PCA50+ rbf SVR	[beauty_baseline_gabors] (beauty_baseline_gabors.py)	0.638
24 colored Gabor filters + PCA50 + rbf SVR	[beauty_baseline_gabors] (beauty_baseline_gabors.py)	0.614
Vgg-ImageNet (pool5+fc6) [2] + linear SVR	[beauty_vgg_imagenet] (beauty_vgg_imagenet.py)	0.804
Vgg-Face (pool5+fc6) [3] + linear SVR	[beauty_vgg_face] (beauty_vgg_face.py)	0.856

Example of prediction

References

[1] Xie, Duorui, Lingyu Liang, Lianwen Jin, Jie Xu, and Mengru Li. "SCUT-FBP: A Benchmark Dataset for Facial Beauty Perception." In Systems, Man, and Cybernetics (SMC), 2015 IEEE International Conference on, pp. 1821-1826. IEEE, 2015.

[2] https://gist.github.com/ksimonyan/fd8800eeb36e276cd6f9

[3] http://www.robots.ox.ac.uk/~vgg/software/vgg_face/

beauty_vision's Issues

error in baseline_random.py

Hello, i encounter the following problem, not sure how to fix it

500 images and 500 labels read
computing features using random filters
Traceback (most recent call last):
File "beauty_baseline_random.py", line 152, in
feat_maps = relu(f2(f1(np.asarray(image_list, dtype='float32').transpose(0,3,1,2))))
File "/home/sukun/.local/lib/python2.7/site-packages/theano/compile/function_module.py", line 917, in call
storage_map=getattr(self.fn, 'storage_map', None))
File "/home/sukun/.local/lib/python2.7/site-packages/theano/gof/link.py", line 325, in raise_with_op
reraise(exc_type, exc_value, exc_trace)
File "/home/sukun/.local/lib/python2.7/site-packages/theano/compile/function_module.py", line 903, in call
self.fn() if output_subset is None else
RuntimeError: BaseCorrMM: Failed to allocate output of 500 x 24 x 216 x 216
Apply node that caused the error: CorrMM{valid, (1, 1), (1, 1), 1 False}(input, Subtensor{::, ::, ::int64, ::int64}.0)
Toposort index: 1
Inputs types: [TensorType(float64, 4D), TensorType(float64, 4D)]
Inputs shapes: [(500, 3, 224, 224), (24, 3, 9, 9)]
Inputs strides: [(1204224, 8, 5376, 24), (1944, 648, -72, -8)]
Inputs values: ['not shown', 'not shown']
Outputs clients: [['output']]
Backtrace when the node is created(use Theano flag traceback.limit=N to make it longer):
File "beauty_baseline_random.py", line 150, in
f1 = theano.function([input], conv2d(input, W))
HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node.