On the Impact of Resolution in Fashion Attribute-based Re- identification

In the realm of computer vision and intelligent surveillance, person re-identification (ReID) plays a pivotal role in matching individuals across non-overlapping multi-camera environments. This study investigates the influence of image resolution on ReID performance, particularly within the context of fashion attribute-based ReID. Fashion attribute-based ReID presents a promising alternative to traditional methods by eliminating the need for facial data, thus addressing privacy concerns and expanding its applicability. Through rigorous and extensive experiments, the relationship between image resolution, model parameters, and instance normalization on ReID are examined for various ReID models. Our findings indicate that ReID performance remains stable for image heights ranging from 150 pixels to 70 pixels, with a notable decline below 70 pixels due to the loss of critical features perceptible to the human eye. This threshold represents a balance point where further resolution enhancement may yield diminishing returns. Additionally, we observed that Rank-1 accuracy is particularly sensitive to changes in image resolution, while higher rank accuracy exhibits greater resilience. Notably, the model's adaptability to half-sized parameters suggests the possibility of flexible adjustments without significant performance degradation. This study underscores the importance of considering image resolution in fashion attribute-based ReID, offering insights into optimizing model parameters and image resolution. Our research contributes valuable knowledge for enhancing ReID in intelligent surveillance, with applications spanning anomaly detection, crowd disaster prevention, and beyond.