TODO : more detail
-> huhm, this is nonsense anyway
Read the slide here : https://drive.google.com/drive/folders/1sTjR75hgDDML2owb9erRdnKVLRvuluo4 https://github.com/timy90022/One-Shot-Object-Detection
one-shot learning : Matching Networks Input: a. anchor : base image b. other test image One-Shot : object detection Output : which image is match
One-Shot Instance Segmentation Given a query image and a reference image showing an object of a novel category, we seek to detect and segment all instances of the corresponding category (in the image above ‘person’ on the left, ‘car’ on the right). https://github.com/bethgelab/siamese-mask-rcnn : 2018 https://github.com/timy90022/One-Shot-Object-Detection : 2019
-> very promising , but it's hard to implemented
Found the following 2 promissing approach
Let’s compare the 1. and 2.
-> We will use 2 for this task
apple has very great turi examples at: https://apple.github.io/turicreate/docs/userguide/one_shot_object_detection/ -> if we need to re-train an one-shot-object-detection then this is a good resource -> don’t know which algorithm are they using
N-shot learning = N sample for training few-shot learning : few = 1~5 -> zero shot learning -> one shot learning -> few-shot learning
training set = support set -> K classes, N examples per class testing set = query set -> K classes, a lot of example per class
-> reduce the size of training set as small as possible -> don’t know why they changed the name
zero-shot learning : do the task on unseed object one-shot learning : do the task while using only 1 samples per class to learn -> take the advantage of siamese (metrics learning) few-shot learning : do the task while using only few samples per class to learn
How to measure performance: n-shot, k-way (n-example, k-class)
-> let’s dive a little bit
Zero-shot learning Learning to Compare: Relation Network for Few-Shot Learning Learning Deep Representations of Fine-Grained Visual Descriptions Improving zero-shot learning by mitigating the hubness problem
One-Shot Learning : siamese network , matching networks Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks One-shot Learning with Memory-Augmented Neural Networks Prototypical Networks for Few-shot Learning
Few-Shot learning -> just extend version of one-shot learning
-> famouse network : prototypical networks -> learn the mapping of an image in metric space (not vector space). -> metrics space : does not has origin -> vector space : has origin
-> now understand what is few-shots learning
—> bat dau khong hieu van de roi, doc lai cho ro nao
https://lilianweng.github.io/lil-log/2018/11/30/meta-learning.html -> very good writing
https://en.wikipedia.org/wiki/Meta_learning_(computer_science) -> a good warm up
meta-learning : learning to learn learning new concepts and skills fast with a few training examples? That’s essentially what meta-learning aims to solve.
A classifier trained on non-cat images can tell whether a given image contains a cat after seeing a handful of cat pictures. A game bot is able to quickly master a new game. A mini robot completes the desired task on an uphill surface during test even through it was only trained in a flat surface environment.
-> common approaches
Optimization-Based What optimization-based approach meta-learning algorithms intend for is to adjust the optimization algorithm so that the model can be good at learning with a few examples.
-> LSTM Meta-Learner -> Temporal Discreteness -> Reptile
Few-shot object recognition became a hot topic recently (from 4 few-shot papers in CVPR18 to around 20 in CVPR19). The usual setup is that you have categories with many examples you can use at training time; then at test time, you are given novel categories (usually 5) with only a few examples per category (usually 1 or 5; called “support-set”) and query images from the same categories. Next, I’ll try to break the few-shot methods to different families. Although these families are not well defined and many methods belong to more than one family. “Older” suggested methods were based on metric learning, where the objective is to learn a mapping from images to embedding space in which images from the same category are closed together and from different categories are far apart. Hoping that it will hold true for unseen categories. Following that, came meta-learning methods. These are models which are conditioned on the current task, so a different classifier is used as a function of the support-set. The idea is to find model hyper-parameters and parameters such that it will be easy to adapt to a new task without over-fitting to the few shots available. In parallel, data augmentation methods are also very popular. The idea is to learn data augmentation so we can generate more examples out of the few examples available. Finally, semantics-based methods are on the rise. It is inspired by zero-shot learning where classification is done based solely on the category name, textual descriptions, or attributes. Those extra semantic ques can also be of help when visual examples are scarce.
One-Shot Learning for Semantic Segmentation -> 2017
Continual Learning (CL) is built on the idea of learning continuously and adaptively about the external world and enabling the autonomous incremental development of ever more complex skills and knowledge.
In the context of Machine Learning it means being able to smoothly update the prediction model to take into account different tasks and data distributions but still being able to re-use and retain useful knowledge and skills during time.
extrinsic : world points -> camera coordinates : extrinsics parameters intrinsic : camera coordinates -> image plane : intrinsics parameters
reason : re-calculate the extrinsics, intrinsic How to evaluate : re-projector error: +/- ??? (in pixel), lower = better examining camera extrinsics : see matlab page for detail View Undistorted Image : see matlab page for detail
A survey of Recent Advances in Texture Representation -> good summary
-> GTOS : Ground Terrain in Outdoor Scenes -> 40 classes, 750 samples each -> https://github.com/jiaxue1993/pytorch-material-classification -> format is a little weird GTOS_MOBILE Ground Terrain in Outdoor Scenes -> use this DTD Describable Textures Dataset -> 47 classes, 120 sample each -> https://www.robots.ox.ac.uk/~vgg/data/dtd/ MINC-2500 Material Recognition in the Wild with the Materials in Context -> 23 material classes, 2500 samples each -> http://opensurfaces.cs.cornell.edu/publications/minc/ FMD/LFMD : Light-Field Material Database -> not 2d data -> ignore
Other dataset : not usefull Other dataset such as Textures Classification dataset is Mere … VisTek : small, no longer maintain alot texture classification dataset : Kylberg Texture Dataset v. 1.0 : gray scale image
texture synthesizability -> texture画像生成方法、画像生成した後に学習隅のモデルで良し悪しを判断 -＞ 約に立たないが、上記のリンクの最後項目にtexture datasetの一覧について説明があります