Comments (2)
from hashnet.
I want to add some of my own thought on this.
The overall goal of image hashing is to map each image to a binary hash code so that ones that are similar shares similar hash codes and ones that are different have different hash codes. The distance is measured in Hamming distance. Binary codes are usually much shorter than the number of classes. This ensures fast retrieval speed since distance between binary codes can be efficiently computed using bitwise xor and smaller hash codes will further boost the speed. So two things that people should care about: (1) image similarity (2) hash code learning
Most papers nowadays are more concerned with improving the retrieval speed and accuracy based on commonly used datasets. This means that (2) are more relevant. (1) is simply based on some common practice in the field. If you look at those datasets, you will notice that they all have class label informations. Even if some paper do not use this directly and formulate it as similarity matrix, you should know that whether images are similar or not will always be computed from class label. For instance in imagenet100 and cifar10, this is simply the class lable. In NUS WIDE and MS COCO, this is whether two images share at least one common class label or not.
And as a side note, the pretrained model used for image hashing will always be from some pretrained Imagenet classification model. You never see people report training from scratch. This is because many simply do not converge. So I would say that nowadays most image hashing models work by projecting one hot vector or multiclass label into a much shorter binary hash codes. This is different from classification because softmax results or some intermediate conv features as used in face recognition are all floating values. You usually calculate the max or the Euclidean distance. Results from them can not be directly mapped to binary codes without fine-tune with image hashing algorithm.
With that being said, if you are interested in (1), you can check out papers like context embedding network or contextual visual similarity. If you don't care about it, then feel free to use any class label information you want if it gives you good results. Many papers do use it to boost their performance and some even strictly use it.
from hashnet.
Related Issues (20)
- Problem with MAP@k
- 你好,非常感谢分享的NUS数据集,但是下载解压发现数据的图片比原始要少 HOT 2
- Why take two batches as input HOT 4
- the mAP on cifar10 is only 0.30 HOT 5
- the scale tanh is useless HOT 1
- a question about training set HOT 1
- The mAP of ImageNet is different from that on paper HOT 7
- How did you generate the "train.txt" of MS-COCO? HOT 1
- The pretrained model HOT 1
- nuswide dataset HOT 1
- @bfan @caozhangjie HOT 1
- question on nuswide dataset
- Pytorch with Resnet50 HOT 6
- HashNet on CUB200 HOT 1
- Anybody willing to share a pretrained model (from ImageNet or CoCo or similar)?
- AlexNet backbone result (COCO) HOT 3
- False sampling of data HOT 1
- test.py for CIFAR HOT 1
- number of training images for imagenet HOT 1
- question on nuswide81 dataset HOT 1
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from hashnet.