Convolutional Neural Network Training for Person Re-Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current person re-identification methods face challenges such as varying poses, scales, and detection errors across different camera views, leading to complex and cumbersome models that do not effectively outperform state-of-the-art techniques.
Innovation Solution
A method for training a convolutional neural network using triplet loss with a three-stream Siamese architecture, where images are processed sequentially with increasing difficulty, and hard triplet mining is employed to select and retrain on triplets with the largest losses, producing compact yet powerful global image representations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If additional modules are introduced to address scale and pose variations, then robustness to pose and scale changes is improved, but device complexity increases
Solution Approach 1:
The patent changes the parameter of image representation from local features to global representations, and modifies the training objective from detection-based to retrieval-based. This allows the model to handle pose and scale variations through learned global embeddings without adding explicit alignment modules, thus improving robustness while maintaining simplicity.
Solution Approach 2:
Instead of adding modules to align body parts as conventional methods do, the patent inverts the approach by using global image representations and learning retrieval directly from whole images. This inversion eliminates the need for explicit pose estimation and alignment modules, reducing complexity while maintaining robustness to pose variations.
2Measurement precision
If sophisticated learning objectives are used to improve generic deep image representations, then re-identification performance is improved, but device complexity increases
Solution Approach 1:
The patent changes the learning objective from complex multi-task losses (combining classification, ranking, and metric learning) to a simpler retrieval-based objective using global image representations. This parameter change in the learning goal achieves competitive performance while avoiding the complexity of sophisticated multi-objective training schemes.
3Measurement precision
If task-specific image representations are designed, then re-identification accuracy is improved, but device complexity increases
Solution Approach 1:
The patent uses universal global image representations that can be applied to various re-identification scenarios without task-specific modifications. The same global representation framework works across different datasets and conditions, achieving task-specific accuracy through general-purpose global features rather than complex task-specific architectures.
4Measurement precision
If pre-processing stages with person detection algorithms are used, then localization accuracy is improved, but reliability decreases due to detection errors
Solution Approach 1:
The patent extracts the re-identification task from the detection-based pipeline and formulates it as direct image retrieval. By removing the detection pre-processing stage and working directly with global image representations, the system avoids propagation of detection errors while maintaining the ability to handle truncated or misaligned persons through robust global features.
Data Source
AI summary
A method for re-identification of a subject in an image by pre-training a convolutional neural network to recognize individuals within a closed set of possible identifications and further pre-training the convolutional neural network using classification loss; training the pre-trained convolutional neural network by sequentially processing a plurality of triplet of images, each triplet containing a query image degraded by adding random noise to a region of the query image, a positive image corresponding to an image of a same subject as in the query image, and a negative image corresponding to an image of a different subject as in the query image by (a) ranking the triplets by the triplet loss computed, (b) selecting a subset of triplets among the plurality of triplets, and (c) retraining the pre-trained convolutional neural network on each of the triplets of the subset of triplets.