Pedestrian re-identification method based on IBN-Net and negative class center triplet loss
Abstract
In light of the impact of camera background, lighting, resolution, or occlusion changes on pedestrian re-identification, this makes it difficult for the person re-identification models to extract discriminative pedestrian features. This paper presents an innovative model for person re-identification. At the level of feature extraction, IBN-Net50-a serves as the backbone network for extracting features unaffected by appearance variations such as image color, brightness, and background. A Coordinate Attention (CA) mechanism module is integrated into the backbone network to suppress irrelevant features and enhance discriminative features. Simultaneously, the global average pooling at the end of the backbone network is replaced with SoftPool to reduce feature loss and retain more fine-grained features. At the loss function level, a negative class center triplet loss is proposed, which relaxes the constraints between anchor samples and negative samples, reducing the impact of outliers on the model training process. Subsequently, the model is trained using joint cross-entropy loss. Finally, The experiments are carried out on two widely used datasets, Market1501 and DukeMTMC-ReID, The results illustrate the efficacy of the suggested approach.
Related Papers
- → Review on Microscopic Pedestrian Simulation Model(2016)53 cited
- THE POSITION OF THE PEDESTRIAN ON CROSSINGS LOCATIONS. MAIN REPORT AND APPENDICES REPORT(1996)
- Analysis on Influence of Design of Pedestrian Bridge to Pedestrian Traffic(2013)
- → CITY PEDESTRIAN BRIDGES(2020)
- → Effective application condition of exclusive pedestrian phases(2021)