Continual Learning of Object Instances

Feature Estimations Based Correlation Distillation for Incremental Image Retrieval

“…Due to its effectiveness and flexibility, the regularization methods have been widely explored for tasks such as image classification [2][3] [6], but are less-explored for image retrieval. Recently, Parshotam et al [7] regularize the representations via a normalized cross-entropy loss, training with metric learning for vehicle identification and retrieval. Chen et al [8] propose regularizing both the representations and probabilities by combining a maximum mean discrepancy loss and a knowledge distillation loss [9] for fine-grained image retrieval (FGIR) [10].…”

“…To these methods, we further train a classifier on the top of the embedding net so that these two methods can be reproduced correctly. NCE loss [7] regularizes the inner product of an anchor-positive feature pair In terms of the plasticity factor λ 2 , we tune this factor for four methods in incremental FGIR until we get their optimal performance. As a result, the corresponding plasticity factors are tuned as 8000, 0.2, 10, and 0.1, respectively.…”