Cross-domain pedestrian re-identification model based on domain invariant features and method thereof
A pedestrian re-identification, domain-invariant technology, applied in the field of pedestrian re-identification, can solve the problem of time-consuming and labor-intensive data collection in the target domain, and the inability to collect data in the target domain, so as to enhance the cross-domain generalization ability and improve the cross-domain generalization. Ability, the effect of eliminating style differences
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0055] The present invention designs a cross-domain pedestrian re-identification model based on domain-invariant features. The cross-domain pedestrian re-identification model is set after the residual module of the domain-invariant feature extraction network, such as figure 2 As shown, it includes a recovery feature module for obtaining recovery features, a feature enhancement module for obtaining discriminative features, and a feature stacker for superimposing recovery features and discriminative features to obtain complete output features
[0056] The recovery feature module is provided with:
[0057] The instance normalization module IN is used to normalize the input original features to obtain the normalized features of the instance;
[0058] Feature Residual Calculator It is used to calculate the residual between the input original feature and the feature normalized by the instance to obtain the residual feature;
[0059] The first attention mechanism module (includ...
Embodiment 2
[0065] This embodiment is further optimized on the basis of the above-mentioned embodiments. The cross-domain pedestrian re-identification method based on domain-invariant features is implemented by using the above-mentioned cross-domain pedestrian re-identification model based on domain-invariant features, as shown in figure 2 shown, including the following steps:
[0066] 1) The input original feature reduces the domain difference between sample features through the instance normalization module IN, and obtains the feature after instance normalization;
[0067] 2) Utilize the feature residual calculator Perform residual calculation on the input original features and the normalized features of the instance to obtain the residual features;
[0068] 3) Using the residual feature first attention mechanism module (including channel attention module CA1 and spatial attention module SA1) to adaptively extract features related to pedestrian identity information based on channel a...
Embodiment 3
[0075] This embodiment is further optimized on the basis of the foregoing embodiments, and the same parts as the foregoing technical solutions will not be repeated here, such as figure 2 As shown, further in order to better realize the cross-domain pedestrian re-identification method based on domain invariant features described in the present invention, the original feature of the input is set to x, and x∈R b×c×h×w , where b, c, h, and w represent the batch size, the number of channels, and the height and width of the feature map, respectively, R b×c×h×w is a b×c×h×w dimensional matrix, then the complete output feature after the attention and style normalization module ASN (that is, the cross-domain person re-identification model) is y∈R b×c×h×w ;
[0076] The normalized feature of the example is set to x 1 , in the step 1), the input original features are obtained by the instance normalization module IN using the following formula to obtain the features after instance norm...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More