Deep Cosine Metric Learning for Person Re-identification

Abstract

Metric learning aims to construct an embedding where two extracted features corresponding to the same identity are likely to be closer than features from different identities. This paper presents a method for learning such a feature space where the cosine similarity is effectively optimized through a simple re-parametrization of the conventional softmax classification regime. At test time, the final classification layer can be stripped from the network to facilitate nearest neighbor queries on unseen individuals using the cosine similarity metric. This approach presents a simple alternative to direct metric learning objectives such as siamese networks that have required sophisticated pair or triplet sampling strategies in the past. The method is evaluated on two large-scale pedestrian re-identification datasets where competitive results are achieved overall. In particular, we achieve better generalization on the test set compared to a network trained with triplet loss.

GAN-Based Pose-Aware Regulation for Video-Based Person Re-Identification

“…With regards to the baseline [51], on MARS, our PDSR technique allows an overall improvement of +3.2% and +4% respectively for the rank-1 accuracy and the mAP Table 3. This improvement is due in part to the weighted fusion strategy WF (+1.6% and +2.9% over the baseline, respectively, for rank 1 accuracy and mAP) and for the remaining part to the WPR technique (+1.6% and +1.1% for rank-1 accuracy and mAP).…”

“…The importance of accounting for the pose/viewpoint invariance problem, in person re-id, has been amply proved by many works. A popular approach is metric learning [61,58,72,22,51,48] where a similarity metric is learned in the space of the video-level feature vectors expressing different views, aiming to increase the intra-class compactness and the inter-class distance of the identities. A different stream of research, complementary to metric learning, tackles the viewpoint problem by focusing on designing/learning more robust feature representations, for example, exploiting the temporal aggregation of multiple frame-level features maps [30] or performing spatial fusion/concatenation of global/local features [56,4].…”

“…This is why it lends itself particularly well to be combined with the existing video-based feature extraction techniques for producing even more discriminative embeddings. In this paper, we follow this combined approach for complementing the incomplete pose information of video-sequences and refer to [38] for the GAN architecture and to [51] for the feature extraction CNN.…”

“…Residual learning-based CNN architecture[51] used for feature extraction. The patch size for the convolutional, pooling and residual layers is 3 × 3.a training set and a testing set, each one with 702 IDs.…”

Aerial Multi-Object Tracking by Detection Using Deep Association Networks

“…• For the tracking model, we train the deep association network [46] on the object hypotheses generated from the detection module and feed it to the the deep sort algorithm [47] for tracking.…”