부산대학교 도서관

This paper introduces the Temporal Transform Network based on Scale Sequences (TTS) for cloth-changing person re-identification in video datasets. The TTS network is designed to capture multi-scale temporal cues within video sequences. It accomplishes this by initially modeling short-term temporal cues between adjacent frames, followed by capturing long-term relationships between non-consecutive frames. In more detail, short-term temporal cues are modeled through parallel inflated convolutions with different time dilation rates, enabling the representation of pedestrian movement and appearance dynamics. Long-term relationships are effectively captured using a temporal self-attention model, mitigating challenges such as occlusion and noise within the video sequence. The TTS network outperforms existing methods across cloth-changing video ReID datasets such as CCVID. For instance, under general settings, our approach exhibits a 1.1% improvement in top-1 accuracy and a corresponding 1.1% increase in mAP compared to the baseline. In cloth-changing settings, we observe a 0.2% enhancement in top-1 accuracy and a notable 1.3% increase in mAP relative to the baseline.