one General information
Pedestrian re-recognition can be simplified into such a process: feature extraction of the original image data, and then feeding it into the objective function for optimization. This process involves three important parts: data, features and objective functions. Therefore, the REID method is divided into three categories according to the research focus. They are research based on data, research based on feature and research based on objective function. \
We define that the purpose of pedestrian re-recognition is to make the performance as good as possible when testing on a data set. (Rank 1, mAP)
The categories are as follows:
1. Data: refers to images in the real sense. Classification is based on the research focus on data preparation
2 features: depth features generated after CNN are classified based on the research focus on the expression of features
3. Objective function: refers to the part that provides feedback information for updating network parameters. Classification is based on the research focus on feedback information.
\
two Classification of instructions
1 Data:
1.1 Data Generation:
1.1.1 expansion in existing data sets [1,2,3]
1.1.2 expanding the target dataset with other datasets [4,5,6]
1.2 Auxiliary Data
1.2.1 Select some data [7]
1.2.2 all use [8,9]
2. Characteristics of the
2.1 Local feature extraction
2.1.1 Hard Alignment [10]
2.1.2 adaptive alignment [11,12,13]
2.2 Global feature extraction
2.2.1 global characteristics of a single sample [14,15,16]
2.2.2 global characteristics of diverse texts [17,18,19]
3. Objective function:
3.1 Traditional objective function
3.1.1 Selection of objective Function [20]
3.1.2 Dynamic change of weight of objective Function [21]
3.2 Newly defined objective function
3.2.1 Improvement in form [22]
3.2.2 improvement from comparison of input data [23,24,25]
1 Data:
1.1 Data Generation:
1.1.1 Expansion in existing data sets
1.1.2 Expand the target dataset with other datasets
1.2 Auxiliary Data
1.2.1 Select some data
1.2.2 All use
Classify according to data, according to whether the research focus is on data preparation. Our REID aims to train a network to perform well on the test set. The training data set greatly affects the results. In order to achieve good test results on a certain target data set, we can use GAN network to generate more artificial image data, or use data from other data sets to help improve the test results on the target data set. Therefore, the subcategories of data are divided into two categories: generating data (generating non-existent images) and auxiliary data (utilizing existing real data).
In 1.1 data generation, there are two branches, namely, expanding existing data sets and expanding target data sets by using other data sets:
1.1.1 To expand the existing data set, the samples in the own data set are increased, and the generated data are all clearly labeled
1.1.2 Expanding oneself with other data means migrating and generating the data in other data sets to the target data set and generating the data with clear labels.
In 1.2 Auxiliary Data, there are two branches: select partial data in a data set and use all data in a data set.
1.2.1 Selecting some data refers to the process of training on the target data set, which requires some data from other data sets as reference proxy pictures.
1.2.2 All-use means that when training on the target data set, images from other data sets are used. The more other auxiliary data, the better the effect
\
\
2. Characteristics of the
2.1 Local feature extraction
2.1.1 hard alignment
2.1.2 Adaptive alignment
2.2 Global feature extraction
2.2.1 Global characteristics of a single sample
2.2.2 Global characteristics of multiple texts
\
REID is essentially a classification problem, which requires the depth features of pictures as the basis of classification. When designing CNN network to extract depth features of images, global features can be extracted from the perspective of the whole image as a representation. Local features can also be extracted from the local position of the image as a representation. So feature is divided into two branches, local feature extraction and global feature extraction respectively.
In 2.1 local feature feature extraction, it is divided into two branches, hard alignment and adaptive alignment respectively.
2.1.1 Hard alignment means that when extracting local features of an image, the selection of local areas is fixed directly from the Angle of length and width
2.1.2 Adaptive alignment refers to locating different parts according to different conditions, and then extracting the features at this position.
\
In 2.2 global feature extraction, there are two branches, namely, global feature of single sample and global feature of multiple texts.
2.2.1 Global feature of a single sample refers to the process of feature extraction. When extracting a certain image, only focus on this image.
2.2.2 Global features of multiple texts refers to that when extracting the features of a picture, we should not only focus on its own features, but also compare and match with other existing features. There are many articles about memory bank, all of which focus on this research.
\
3. Objective function:
3.1 Traditional objective function
3.1.1 Selection of objective function
3.1.2 Dynamic change of weight of objective function
3.2 Newly defined objective function
3.2.1 Improve from the form
3.2.2 Improve from the comparison of input data
\
Objective function, provide feedback information for network training. Studying the REID problem in terms of objective functions is equivalent to focusing on how to use existing information. The research of objective function is divided into two branches, namely traditional objective function and newly defined objective function.
3.1 Traditional objective function refers to the commonly used objective function of classification problems, including cross entropy, etc. This part is divided into two branches, which are the selection of objective function and the dynamic change of weight of objective function.
3.1.1 Selection of objective function. In REID training, cross entropy is often used to calculate the classification error and then optimize the network. In addition to this cross entropy, other objective functions are often added to improve the training effect.
3.1.2 Dynamic change of the weight of the objective function means that when the objective function is determined, the weight of the function will change with the change of the number of training rounds and some values.
3.2 Defining new objective functions refers to adding some novel objective functions in the training process to improve the expression ability of features. Unsupervised learning is often used.
\
- Corresponding to the literature
- Joint Discriminative and Generative Learning for Person Re-identification
- Progressive Pose Attention Transfer for Person Image Generation
- Unsupervised Person Image Generation with Semantic Parsing Transformation
- Image-Image Domain Adaptation with Preserved Self-Similarity and Domain-Dissimilarity for Person Re-identification
- Person Transfer GAN to Bridge Domain Gap for Person Re-Identification
- SBSGAN: Suppression of Inter-Domain Background Shift for Person Re-Identification source
- Unsupervised Person Re-identification by Soft Multilabel Learning
- Distilled Person Re-identification: Towards a More Scalable System
- Self-Similarity Grouping: A Simple Unsupervised Cross Domain Adaptation Approach for Person Re-Identification
- Beyond Part Models: Person Retrieval with Refined Part Pooling (and A Strong Convolutional Baseline)
- Densely Semantically Aligned Person Re-Identification
- Attention Towards Rich Feature Discovery with Class Activation Maps Augmentation for Person Re-Identification
- AANet: Attribute Attentio Network for Person Re-Identification
- Spectral Feature Transformation for Person Re-Identification
- Adaptive Transfer Network for Cross-Domain Person Re-Identification
- View Confusion Feature Learning for Person Re-Identification
- Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-identification
- Generalizable Person Re-identification by Domain-Invariant Mapping Network
- Unsupervised Person Re-identification by Soft Multilabel Learning
- In Defense of the Triplet Loss for Person Re-Identification
- Pyramidal Person Re-IDentification via Multi-Loss Dynamic Training
- Unsupervised Person Re-identification by Soft Multilabel Learning
- self-Similarity Grouping: A Simple Unsupervised Cross Domain Adaptation Approach for Person Re-Identification
- Maximum-value Perfect Matching for Mining Hard Sample
- A Novel Unsupervised Camera-Aware Domain Adaptation Framework for Person Re-Identification
\
\
\
\
\