CVPR Daily - Tuesday

The primary purpose of generalized few-shot object detection tasks is to train a detector to detect the instances from base classes, which have a lot of training data, and novel classes, which have limited training data. TFA is a widely-used framework in generalized few-shot object detection tasks. Models are pre-trained on data from base classes and then fine- tuned on a union of base and novel classes. This fine-tuning is called few- shot adaptation . An essential part of this second step is aggressively down- sampling the base training set to achieve a balanced training set among the base and novel classes. “ Few-shot adaptation is key to the success of this framework, but it also causes a problem, ” Jiawei tells us. “ By aggressively down-sampling the training data of the base classes, it sacrifices the detection precision . With limited training data, the model will overfit to those few samples of base classes. There is always a trade-off. We achieve good performance on the few-shot novel classes at the expense of the detection precision on the base classes . ” You can only perform a good few-shot novel adaptation or detect the instances of novel classes correctly when you have well-separated classifier weights . Down-sampling is crucial because only when the model is trained on a balanced data set will the classifier weights be maximally separated. Jiawei Ma is a fourth-year PhD candidate at Columbia University, advised by Shih-Fu Chang. He proposes a new training framework for generalized few-shot object detection tasks and speaks to us ahead of his poster this morning. DiGeo: Discriminative Geometry-Aware Learning for Generalized Few-Shot Object Detection 12 DAILY CVPR Tuesday Poster Presentation