Constraint Patch Matching for Faster Person Re-identification

In person re-identification, it is of great importance toextract very discriminative features in order to distinguishbetween images of different persons captured in differentcamera views. Features can be extracted globally from theentire or parts of the image, or locally from patches. Inaddition, accuracy can be increased by utilizing supervisedlearning before matching. Matching can either be done us-ing global feature descriptors or local patch features withthe latter being more computationally heavy due to the num-ber of patch pairs to match. Horizontal constraints aretherefore usually added to only match patches at same hor-izontal location. As an extension, we have developed an al-gorithm to add vertical constraints to different body parts,to increase accuracy and decrease processing time. Theconstraints are applied to the CVPDL system by Liet al.[13] in order to compare accuracy with the use of a fore-ground mask and processing time when only adding hori-zontal constraints. We refer to our constrained CVPDL asC-CVPDL. Experiments conducted on two datasets, VIPeRand CUHK01, show C-CVPDL to achieve similar rank-1accuracy on VIPeR while improving rank-1 accuracy by3.83% for CUHK01 compared to CVPDL. Furthermore,experimental results on CUHK03 show a rank-1 accuracy52.05%, being comparable to state-of-the-art CNN’s andbeating other patch matching systems. Finally, timings forVIPeR and CUHK01 show our constraints to decrease timeby 32.77% and 37%, respectively, while only taking up to36ms to compute per person.