CAD: Scale Invariant Framework for Real-Time Object Detection

Abstract

Real-time detection frameworks that typically utilize end-to-end networks to scan the entire vision range, have shown potential effectiveness in object detection. However, compared to more accurate but time-consuming frameworks, detection accuracy of existing real-time networks are still left far behind. Towards this end, this work proposes a novel CAD framework to improve detection accuracy while preserving the real-time speed. Moreover, to enhance the generalization ability of the proposed framework, we introduce maxout [1] to approximate the correlation between image pixels and network predictions. In addition, the non-maximum weighted (NMW) [2] is employed to eliminate the redundant bounding boxes that are considered as repetitive detections for the same objects. Extensive experiments are conducted on two detection benchmarks to demonstrate that the proposed framework achieves state-of-the-art performance.