Random removal of textures during data-augmentation or use Stylized-ImageNet +2.9 mAP@0.5

[AlexeyAB]

• Paper: https://openreview.net/forum?id=Bygh9j09KX

• PDF: https://openreview.net/pdf?id=Bygh9j09KX

• Code to create Stylized-ImageNet, a stylized version of standard ImageNet: https://github.com/rgeirhos/Stylized-ImageNet

• re-trained models, data, code & materials from the paper: https://github.com/rgeirhos/texture-vs-shape

---

Random removal of textures during data-augmentation - something like Stylized-ImageNet, just we remove textures instead of changing then. Try to use something like Blur-then-sharpen: https://docs.opencv.org/master/d1/d10/classcv_1_1MatExpr.html#details

Mat blurred; double sigma = 1, threshold = 5, amount = 1;
GaussianBlur(img, blurred, Size(), sigma, sigma);
Mat lowContrastMask = abs(img - blurred) < threshold;
Mat sharpened = img*(1+amount) + blurred*(-amount);
img.copyTo(sharpened, lowContrastMask);

• Or super-resolution: https://docs.opencv.org/2.4/modules/superres/doc/super_resolution.html
• Or something else: https://stackoverflow.com/questions/4993082/how-to-sharpen-an-image-in-opencv

Based on articles:

• Approximating CNNs with bag-of-local features models works surprisingly well on ImageNet https://arxiv.org/pdf/1904.00760.pdf
• ImageNet-trained CNNs are biased towards texture https://arxiv.org/pdf/1811.12231.pdf
• Texture Synthesis Using Convolutional Neural Networks https://arxiv.org/abs/1505.07376

[LukeAI]

Maybe more useful repo: https://github.com/bethgelab/stylize-datasets A generalisation of the Imagenet repo above for styling arbitrary datasets.

[AlexeyAB]

@LukeAI Why is it better?

[LukeAI]

I haven't used it but it works with arbitrary datasets, not just Imagenet.

[AlexeyAB]

Implemented bilateralFilter-bluring - it keeps details, but removes textures. It randomly alternates between blurry and non-blurry images, so that the textures will not be completely lost: https://github.com/AlexeyAB/darknet/commit/142fcdeb1e53ec78ec35d98503726075bd721a9b

You can use (only if Darknet is compiled with OpenCV):

[net]
blur=1

So will be applied cv::bilateralFilter(src, dst, ksize, 75, 75); - by default ksize=17

• if blur=0 - blurring will not be used
• if blur=1 - will be used ksize=17 - randomly the image will be blurry or not
• if blur > 1 - will be used ksize=blur - randomly the image will be blurry or not

---

---

---

[924175302]

@AlexeyAB Hello, I want to experiment with this effect on my data set. How can I output a randomly deleted texture image just like you? Because I want to compare the experimental results on my data with the original image and the randomly deleted texture image, what are the specific implementation steps, I hope you can answer

[AlexeyAB]

@924175302 Hi,

Set in cfg-file and train your model:

[net]
blur = 10