날아가는 개발자

분류 전체보기 검색 결과

해당 글 124건

Graph Convolutional Network (GCN)

Graph 1) 개요 node와 edge로 이루어진 구조를 그래프라고 한다. 이를 활용하여 데이터 간의 관계성이 크고 중요한 데이터를 효과적으로 나타낼 수 있다. 페이스북의 친구관계, 분자구조, 바이러스 감염자 간의 관계 등이 대표적인 예이다. edge는 단순히 연결하는 것에 그치지 않고 방향성(directed, undirected)과 가중치(weighted, unweighted)를 활용하여 추가적인 정보를 가지기도 한다. 2) 그래프 데이터의 구조 그래프 데이터는 vertex (node) set와 edge set로 구성되며 다음의 총 4가지 요소들로 나타낼 수 있다. - node-feature matrix (n x f) : 각 노드에 해당되는 feature 정보를 가지고 있는 matrix로 featur..

AI/Deep Learning 2021. 4. 12. 18:19

Camera

Camera 1) Pin-hole camera - Pin-hole camera model is a widely used camera model in computer vision. - It collects light through a small hole to the inside of dark box or room. - Light passes through a single point, the camera center, C, before it is projected onto an image plane. 2) Lens camera - Lenses map bundles of rays from points on the scene to the sensor. 3) Common feature - only use cent..

AI/Computer Vision 2021. 4. 8. 15:43

RANSAC

Image correspondence pipeline (Panoramas) 1) feature point detection - detect corners using the Harris corner detector. 2) feature point desription - describe features using the Multi-scale oriented patch descriptor. 3) feature matching 4) combine the images together RANSAC; RANdom SAmple Consensus 1) 개요 - Iterative method to fit models to data that can contain outliers. - RANSAC can hopefully p..

AI/Computer Vision 2021. 4. 6. 20:06

Geometric Transformation

2D Transformation 2D linear transformations 1) 개요 - change domain of image 2) Properties of linear transformations - Origin maps to origin - Lines map to lines - Parallel lines remain parallel - Ratios are preserved - Closed under composition 3) parametric transformation $$\begin{bmatrix} x' \\ y' \end{bmatrix} = T \begin{bmatrix} x \\ y \end{bmatrix}$$ - Transformation T is a coordinate-changin..

AI/Computer Vision 2021. 4. 4. 15:50

RNN, LSTM, GRU

RNN 1) sequence data - sequence data is closely connected itself by an order. - MLP or CNN cannot handle sequence data. 2) advantages - Can process any length input - Computation for step t can (in theory) use information from many steps back - Model size doesn't increase for longer input - Same weights applied on every timestep, so there is symmetry in how inputs are processed. 3) disadvantages..

AI/Deep Learning 2021. 4. 2. 15:51

Hough Transform

Hough Transform 1) 개요 - Generic framework for detecting a parametric model. - Edges don't have to be connected. - Lines can be occluded. - Edges vote for the possible models. - Robust to noise. 2) 원리 - Points(x, y) can be explained to parameter space m and b. - Edges can be interpreted as many points in the image. If these points are quantized parameter space, there would be many points of inter..

AI/Computer Vision 2021. 4. 1. 20:06

Image Resampling (Image pyramids)

Image downsampling 1) 개요 - Just deleting even rows and even columns occurs bad condition of images. 2) Aliasing - Occurs when sampling rate is not high enough to capture the amount of detail in the image. - Undersampling can disguise a signal as one of a lower frequency. 3) Anti-aliasing - To avoid aliasing, sampling rate >= 2 * max frequency in the image (Nyquist rate) - Apply a smoothing filte..

AI/Computer Vision 2021. 3. 30. 23:18

Deep residual learning for image recognition; ResNet (2015)

ResNet 1) 개요 - very deep networks using residual connections - 152 layers - ILSVRC 15 winner - Between 20 and 56 layers 'plain' convolutional networks, the deeper network has higher training error and test error.> - Try to resolve degradation problem with deeper network which isn't caused by overfitting. Residual Block - Use network layers to fit a residual mapping instead of directly trying to ..

AI/Paper 2021. 3. 29. 19:23

Going deeper with convolutions; GoogLeNet (2014)

GoogLeNet 1) 개요 - deeper networks, with computational efficiency - 14 ILSVRC winner - 22 layers - Only 5 million parameters (12x fewer parameters than the AlexNet) - Efficient "inception" module - NO FC layers Inception module 1) Naive version - design a local network topology (network within a network) and then stack these modues on top of each other. - apply parallel filter operations on the i..

AI/Paper 2021. 3. 29. 14:57

Very deep convolutional networks for large-scale image recognition; VGGNet (2015)

VGGNet 1) 개요 - Investigate the effect of the convolutional network depth on its accuracy. - 2nd place at ILSVRC (1st : GoogLeNet). But, it is used alot more than GoogLeNet due to simple structure and generalize well to other dataset. - Similar training procedure as AlexNet. 2) Depth (small filters, deeper networks) - The depth of networks were shallow before 2014 ILSVRC. (less than 8 layers) - V..

AI/Paper 2021. 3. 29. 13:23