Everything You Wanted to Know about Deep Learning for Computer Vision but Were Afraid to Ask

Abstract

Abstract-Deep Learning methods are currently the stateof-

Unsupervised Representation Learning Using Convolutional and Stacked Auto-Encoders: A Domain and Cross-Domain Feature Space Analysis

“…computed for all training examples be the mean squared error, then the undercomplete AE is able to learn the same subspace as the PCA (Principal Component Analysis), i.e. the principal component subspace of the training data [2]. Because of this type of behaviour AEs were often employed for dimensionality reduction.…”

“…For image data, Convolutional Neural Networks (CNNs) with multiple layers were found to be particularly adequate. After being trained for image classification tasks, those network models were shown to be good extractors of low-level (shapes, colour blobs and edges) at the initial layers, and highlevel features (textures and semantics) at deeper layers [2]. However, deep networks are difficult to train from scratch, requiring a large number of annotated examples in order to ensure learning, due to their high shattering coefficient [3].…”

“…In this paper we explore stacked and convolutional autoencoders as alternative methods for unsupervised feature learning. Auto-encoders (AEs) are methods that encode some input into a representation with low dimensionality, known as code, and then, via a decoding module, reconstruct this compact representation to match as best as possible the original input [2] and can be useful in many scenarios, in particular for signal, image and video [10], [11] applications. The space formed by the transformation of the input into the code is often called latent space.…”

Features transfer learning between domains for image and video recognition tasks

“…In both scenarios, TL has the potential to be applied to various problems, such as for example traffic control , facial attribute classification (ZHUANG et al, 2018), video classification (WU et al, 2015), and anomaly detection in surveillance videos (XU et al, 2017). As it is the current standard in computer vision (PONTI et al, 2017), all the aforementioned studies applied Deep Learning (DL) as a tool (see more details in section 2.1), exploring TL via architecture retraining or feature extraction. Because of the hierarchical structure of Deep Neural Networks (DNN), such methods are able to represent both low-level (shapes, borders, and colors) and high-level (texture and semantics) visual features (YOSINSKI et al, 2014;PONTI et al, 2017).…”

“…As it is the current standard in computer vision (PONTI et al, 2017), all the aforementioned studies applied Deep Learning (DL) as a tool (see more details in section 2.1), exploring TL via architecture retraining or feature extraction. Because of the hierarchical structure of Deep Neural Networks (DNN), such methods are able to represent both low-level (shapes, borders, and colors) and high-level (texture and semantics) visual features (YOSINSKI et al, 2014;PONTI et al, 2017). In Convolutional Neural Networks (CNNs), different processing layers can be incorporated, where convolutional, dense, and pooling are the most relevant ones.…”