Relationship Matters: Relation Guided Knowledge Transfer for Incremental Learning of Object Detectors

By Kandan Ramakrishnan, Rameswar Panda, Quanfu Fan, John Henning, Aude Oliva, Rogério Schmidt Feris. 2020

12 citations recorded by the source corpus.

Standard deep learning based object detectors suffer from catastrophic forgetting, which results in performance degradation on old classes as new classes are incrementally added. There has been a few recent methods that attempt to address this problem by minimizing the discrepancy between individual object proposal responses for old classes from the original and the updated networks. Different from these methods, we introduce a novel approach that not only focuses on what knowledge to transfer but also how to effectively transfer for minimizing the effect of catastrophic forgetting in incremental learning of object detectors. Towards this, we first propose a proposal selection mechanism using ground truth objects from the new classes and then a relation guided transfer loss function that aims to preserve the relations of selected proposals between the base network and the new network trained on additional classes. Experiments on three standard datasets demonstrate the efficacy of our proposed approach over state-of-the-art methods.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Jun Zhu
2. Hang Su
3. Liyuan Wang
4. Xiaopeng Hong
5. Xingxing Zhang
6. Xing Wei
7. Xiaoyu Tao
8. Yihong Gong

In Computer Vision and Pattern Recognition

1. Jun Zhu
2. Hang Su
3. Xiaopeng Hong
4. Liyuan Wang
5. Xingxing Zhang
6. Xing Wei
7. Xiaoyu Tao
8. Xinyuan Chang

In Representation & Self-Supervised Learning

1. Liyuan Wang
2. Yihong Gong
3. Jun Zhu
4. Xing Wei
5. Hang Su
6. Songlin Dong
7. Xingxing Zhang
8. Xiaopeng Hong

Citations in the corpus

• A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, methodology): “2, d), which is further validated by a theoretical analysis of the forgetting-generalization trade-off [350] and the PACBayes bound of generalization errors [330], [430].”
• Few-Shot Class-Incremental Learning via Relation Knowledge Distillation (2021, background): “(3) Relation-based knowledge further explores the structures between different layers and data samples (Yim et al. 2017; Park et al. 2019; Ramakrishnan et al. 2020).”
• Multi-task incremental learning for object detection (2020, unknown): “Distillation-based methods are further improved by [15, 36, 49, 53].”