Learning Generalisable Omni-Scale Representations for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Person re-identification (re-ID) requires models that learn discriminative and generalizable features.
• Existing models often struggle with cross-dataset discrepancies and computational efficiency.

Purpose of the Study:

• To develop novel Convolutional Neural Network (CNN) architectures for effective person re-identification.
• To address the challenges of learning discriminative and generalizable features for re-ID models.

Main Methods:

• Introduced the omni-scale network (OSNet) with a novel building block capturing multi-scale features via a unified aggregation gate.
• Incorporated instance normalization (IN) layers to enhance cross-dataset generalizability.
• Employed a differentiable architecture search algorithm to optimize IN layer placement.

Main Results:

• OSNet achieved state-of-the-art performance in the same-dataset setting, outperforming larger models.
• In the cross-dataset setting, OSNet surpassed most unsupervised domain adaptation methods without target data.
• The proposed OSNet is computationally lightweight due to factorized convolutions.

Conclusions:

• OSNet effectively learns omni-scale features for discriminative person re-identification.
• Instance normalization and architecture search significantly improve cross-dataset performance.
• OSNet offers a promising solution for efficient and generalizable person re-identification.