Pedestrian re‐identification via coarse‐to‐fine ranking

Appearance‐based person re‐identification is particularly difficult due to varying lighting conditions and pose variations across camera views. Taking inspiration from image retrieval, in which windowed searching over locations is proven to be more effective, the authors first perform dense local feature matching using graph cuts to properly deal with the pose variation problem. However, the re‐identification problem suffers from far more overlap between feature distributions. In a re‐identification problem, many samples cropped from surveillance videos are heavily contaminated by external factors or internal mechanical noises, making the images from the same pedestrian totally different. These overly difficult samples would significantly degenerate the training performance. To address this problem, a query‐level loss function for ranking is proposed, benefiting from taking into account the training data every query set to decrease the punishment for those morbid samples. The authors further develop a coarse‐to‐fine iterative algorithm, where the update in each iteration is computed by solving a gradient‐based optimisation and update iteration is to refine the training data by adjusting an ‘ Expected rank ’ parameter. The authors present experiments to demonstrate the performance gain of the proposed method over existing template matching and ranking models.