Any-Shot Object Detection

Previous work on novel object detection considers zero or few-shot settings where none few examples of each category are available for training. In real world scenarios, it is less practical to expect that ‘all’ the classes either unseen have few-examples. Here, we propose a more realistic setting termed ‘Any-shot detection’, totally and categories can simultaneously co-occur during inference. Any-shot offers unique challenges compared conventional such as, high imbalance between unseen, seen classes, susceptibility forget base-training while learning distinguishing from background. To address these challenges, unified any-shot model, concurrently learn detect both zero-shot classes. Our core idea use class semantics as prototypes detection, formulation naturally minimizes knowledge forgetting mitigates class-imbalance in label space. Besides, rebalanced loss function emphasizes difficult cases but avoids overfitting allow Without bells whistles, our framework also be used solely Zero-shot Few-shot tasks. We report extensive experiments Pascal VOC MS-COCO datasets approach shown provide significant improvements.