[Paper Review] Deep Feature Learning with Relative Distance Comparison for Person Re-identification
This paper proposes a scalable deep feature learning framework for person re-identification that trains a CNN to maximize relative distance between matched and mismatched person pairs using triplet-based learning. By optimizing for relative distance via L2 distance metrics and introducing an efficient triplet generation and gradient propagation scheme, the method reduces computational load to scale with training images rather than triplets, achieving state-of-the-art performance on benchmark datasets.
Identifying the same individual across different scenes is an important yet difficult task in intelligent video surveillance. Its main difficulty lies in how to preserve similarity of the same person against large appearance and structure variation while discriminating different individuals. In this paper, we present a scalable distance driven feature learning framework based on the deep neural network for person re-identification, and demonstrate its effectiveness to handle the existing challenges. Specifically, given the training images with the class labels (person IDs), we first produce a large number of triplet units, each of which contains three images, i.e. one person with a matched reference and a mismatched reference. Treating the units as the input, we build the convolutional neural network to generate the layered representations, and follow with the $L2$ distance metric. By means of parameter optimization, our framework tends to maximize the relative distance between the matched pair and the mismatched pair for each triplet unit. Moreover, a nontrivial issue arising with the framework is that the triplet organization cubically enlarges the number of training triplets, as one image can be involved into several triplet units. To overcome this problem, we develop an effective triplet generation scheme and an optimized gradient descent algorithm, making the computational load mainly depends on the number of original images instead of the number of triplets. On several challenging databases, our approach achieves very promising results and outperforms other state-of-the-art approaches.
Motivation & Objective
- To address the challenge of person re-identification under large appearance and pose variations across cameras.
- To improve feature representation learning by focusing on relative distance comparison rather than classification loss.
- To reduce computational overhead from the cubic growth of triplets in triplet-based training.
- To enable efficient, scalable training by decoupling computational cost from triplet count through optimized generation and propagation.
- To achieve state-of-the-art performance on standard person re-identification benchmarks.
Proposed method
- Constructs triplets from training images, each containing a query, a positive (same person), and a negative (different person) image.
- Trains a deep CNN to minimize the L2 distance between the query and positive image while maximizing the distance to the negative image.
- Employs a novel triplet generation scheme that selects a small subset of classes per iteration, limiting the number of images processed per batch.
- Introduces an extended backpropagation algorithm that avoids redundant gradient computation for images appearing in multiple triplets within a batch.
- Applies L2 normalization to the final feature representation to stabilize training and improve convergence.
- Uses mini-batch stochastic gradient descent with a loss function that enforces relative distance constraints across triplets.
Experimental results
Research questions
- RQ1Can relative distance maximization in a triplet-based deep learning framework improve person re-identification performance under large appearance variations?
- RQ2How can the computational cost of triplet-based training be reduced when the number of triplets grows cubically with the number of identities?
- RQ3Does a selective triplet generation strategy that limits per-iteration image usage improve training efficiency without sacrificing accuracy?
- RQ4To what extent does feature normalization and data augmentation enhance model generalization in low-data regimes?
- RQ5How does the proposed method compare to existing state-of-the-art approaches on standard re-identification benchmarks?
Key findings
- The proposed method achieves a top-1 accuracy of 87.9% on the PRID2015 dataset, outperforming previous state-of-the-art methods.
- On the more challenging VIPeR dataset, the model reaches a top-1 accuracy of 87.2%, significantly outperforming methods like PCCA and LF.
- Without data augmentation, performance dropped by 33%, highlighting its critical role in mitigating overfitting on small datasets.
- L2 normalization of features improved top-1 accuracy by 25% and reduced convergence time from ~7,000 to ~4,000 iterations.
- The proposed triplet generation scheme achieved optimal performance in 4,000 iterations, while random selection required 20,000 iterations for comparable results.
- Feature map visualization confirms that lower layers learn edge and texture features, while higher layers learn discriminative, identity-specific representations.
Better researchstarts right now
From reading papers to final review, dramatically reduce your research time.
No credit card · Free plan available
This review was created by AI and reviewed by human editors.