Person Re-Identification ANN Adaptation With Support-Set Distillation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for adapting an artificial neural network (ANN) model trained on a source domain to a target domain suffer from catastrophic forgetting and domain shift, especially when data from the target domain is limited and biased, and require large data storage which is impractical or prohibited.
Innovation Solution
A method involving multiple iterations of an adaptation phase that includes constructing a support set of similar source images, using a teacher model for knowledge distillation, and calculating domain shift costs to minimize global costs, thereby updating the ANN model weights to adapt to the target domain without needing large data storage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If online unsupervised domain adaptation is performed with limited target domain data, then the model can adapt to the target domain without gathering large amounts of data, but catastrophic forgetting occurs and the model forgets knowledge acquired previously
Solution Approach 1:
The method constructs a support set from source domain images that are similar to the target domain images before performing domain adaptation. This preliminary construction of relevant source data provides a foundation that guides the adaptation process, enabling the model to adapt to the target domain while maintaining knowledge from the source domain through the similarity-based selection mechanism
Solution Approach 2:
The support set acts as an intermediary between the source domain and target domain. It is constructed from source images that are similar to target images and serves as a bridge during the domain adaptation process, allowing knowledge transfer while preventing catastrophic forgetting by maintaining relevant source domain characteristics
2Quantity of substance
If online unsupervised domain adaptation is performed with limited target domain data, then data storage requirements are reduced, but domain shift occurs since the model accesses a small and possibly biased subset of target domain data
Solution Approach 1:
The support set is constructed in advance from source domain images that are similar to the target domain images. This preliminary action selects source images that will be most useful for domain adaptation, reducing the need to store large amounts of target domain data while minimizing domain shift through similarity-based selection
Solution Approach 2:
The method changes the parameter of data selection from random or uniform sampling to similarity-based sampling. By constructing the support set using a similarity function that measures the relationship between source and target images, the method adapts the data selection parameter to reduce domain shift while maintaining reduced storage requirements
3Reliability
If the model is retrained on the target domain, then the model achieves good performance on the target domain, but it requires collecting and annotating data from the target domain which is not practical or possible
Solution Approach 1:
The support set constructed from source domain images serves as an intermediary that eliminates the need for target domain data collection and annotation. By selecting source images that are similar to target images and using them in the domain adaptation process, the method achieves good target domain performance without requiring practical data collection and annotation efforts
Solution Approach 2:
The method creates a copy of relevant source domain data (the support set) that is similar to the target domain. Instead of collecting and annotating actual target domain data, the method copies and adapts source domain images that have similar characteristics, achieving the same effect without the practical burden of target domain data collection
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a method (100) for adapting an ANN model, previously trained for person re-identification on a source domain (SD), on a target domain, said method (100) comprising several iterations of an adaptation phase (110) comprising: -constructing (120) a support set, by selecting from said source domain (SD) images similar to images in a new set of images received from the target domain (NTSi), -several iterations of a training step (130) comprising the following steps: ▪ determining (140) a rei-id cost on said new set of images (NTSi); ▪ determining (150) a Knowledge Distillation, KD, cost, with respect to a teacher model, on a support set (SSi-i) constructed during a previous iteration of the adaptation phase, ▪ updating (162) the ANN; ▪ updating (164) the teacher model. The invention further relates to a computer program and a device configured to carry out such a method, and to an artificial neural network trained with such a method.