Novel Single Stage Detectors for Object Detection

Most of the recent successful methods in accurate object detection utilized some variants of R-CNN style two stage Convolutional Neural Networks (CNN) in which plausible regions were proposed in the first stage followed by a second stage for decision refinement. These methods are accurate but hard and slow to train. Single stage detection methods, on the other hand, enjoy the high speed of training and the efficiency in deployment. But they have not been as competitive as two stage methods in terms of accuracy such as mAP for high IoU threshold. Recently, Recurrent Rolling Convolution (RRC) architecture, a novel single stage end-to-end object detection network over multi-scale feature maps to construct object classifiers and bounding box regressors, was proposed. The RRC model has achieved state-of-theart performance in some tasks. In our project, we introduce Backward Recurrent Rolling Convolution (BRRC) based on RRC, and show that BRRC is able to produce better results and meanwhile faster than original RRC. We also investigate SSD with more bounding boxes and introduce an encoder-decoder structure, Detection SegNet, for object detection. We evaluate and compare all these models based on IoU scores.