Re-Identification for Multi-Object Tracking Using Triplet Loss
Koung-Suk Ko, Woo Jin Ahn, Geon‐Hee Kim, Myo–Taeg Lim, Tae‐Koo Kang, Dong-Sung Pae · 2021
Assigning a consistent identification(ID) number is a chronic problem in the tracking model. However, recent tracking models lose the ID because it focuses only on the previous frame. This paper constructed a tracking deep learning model using triplet loss to give consistent ID to objects detected while tracking. We also show the best way for pre-processing the input for the triplet-tracking model, which inputs various image sizes. The experimental result of 97.76% accuracy on KITTI shows the effectiveness of our result.