Anchor pruning for object detection

This paper proposes anchor pruning for object detection in one-stage anchor-based detectors. While techniques are widely used to reduce the computational cost of convolutional neural networks, they tend focus on optimizing backbone networks where often most computations are. In this work we demonstrate an additional technique, specifically detection: pruning. With more efficient and a growing trend deploying detectors embedded systems post-processing steps such as non-maximum suppression can be bottleneck, impact anchors head is becoming increasingly important. work, show that many removed without any loss accuracy. retraining, even lead improved Extensive experiments SSD MS COCO made up 44% while simultaneously increasing Further RetinaNet PASCAL VOC general effectiveness our approach. We also introduce `overanchorized' models together with eliminate hyperparameters related initial shape anchors. Code available at https://github.com/Mxbonn/anchor_pruning.