개발로 하는 개발

Depth Adversarial Pertubation : monocular depth estimation에 대한 adversarial attack이라는 용어보다 나오는 내용이 많은 키워드이다. adversarial attack과 관련된 논문을 읽다가 찾게 되었다. 그런데 이 Monodepth와 관련된 adversarial attack은 두 가지로 나뉜다. 첫번째는 Universal attack, 두번째는 Physical attack.1. Universal Adversarial Pertubation- It should be noted that these methods implicitly assume ‘‘digital adversarial attack’’, in which the attacks are imp..

abstract - identify trade-off between robustness and accuracy - empirical하게 연구되었지만, 이 trade-off의 기반을 이루는 theory에 대해서는 알려진게 많지 않다. - prediction error for adversarial examples = natural error + boundary error - differentiable upper bound uniform over all probability distributions and measurable predictors Introduction problem of adversarial defenses -> classifier with high test accuracy on both na..

1. Towards Deep Learning Models Resistant to Adversarial Attacks Adversarial Attack : inputs that are indistinguishable from natural data and yet classified incorrectly by the network -> deel learning model의 inherent(고유의, 타고난) weakness일 수도 Robust Optimization의 관점에서 adversarial robustness of neural network에 접근 - adversarially chosen inputs vs benign inputs : benign input은 classify를 제대로 하는데, 인간 눈으..

Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Network 을 읽고 정리한 내용입니다. - Introduction 기존 object detection methods : - Selective Search (2s/image) : pixel 기준의 유사점 찾기 1. Image segmentation -> initial regions 2. Calculate similarity between neighbouring region pairs 3. Get highest similarity regions and merge 4. Recalculate similarity and add it to the set 5. Iterate 2-4 - Ed..

Image-level prediction : classification -> output : probability vector class $class \times 1$ ( input 관계없이 output size가 정해져 있다) Pixel-level prediction ( input : $224\times224\times3$) : segmentation -> output : probability volume $224 \times 224 \times class$ : denoising -> output : clean(denoised) image $224\times 224 \times 3$ ( input 크기에 비례해 output 크기가 정해진다) - Semantic Segmentation -> pixel (..