Continual Learning of Object Instances
This addresses incremental learning for fine-grained object recognition, but it is incremental as it builds on existing continual learning techniques.
The paper tackles catastrophic forgetting in distinguishing car instances via continual learning, showing that combining Normalised Cross-Entropy regularization and synthetic data transfer reduces forgetting across multiple datasets and methods.
We propose continual instance learning - a method that applies the concept of continual learning to the task of distinguishing instances of the same object category. We specifically focus on the car object, and incrementally learn to distinguish car instances from each other with metric learning. We begin our paper by evaluating current techniques. Establishing that catastrophic forgetting is evident in existing methods, we then propose two remedies. Firstly, we regularise metric learning via Normalised Cross-Entropy. Secondly, we augment existing models with synthetic data transfer. Our extensive experiments on three large-scale datasets, using two different architectures for five different continual learning methods, reveal that Normalised cross-entropy and synthetic transfer leads to less forgetting in existing techniques.