Tiny-RetinaNet: A One-Stage Detector for Real-Time Object Detection

In this paper, we present a new one stage detector for object detection. In order to meet the requirements of real-time detection, we use MobileNetV2-FPN as backbone for feature extraction. The lightweight depthwise separable convolutions can improve the speed of detection and make the model smaller. In order to improve the accuracy of our small detection model, we add Stem Block into backbone and we add SEnet in front of two task-specific subnets. The stem block can reduce the information loss from raw input images. The SEnet can enhance useful features from backbone network and suppress features that are little use to two-specific tasks. Inspired by RetinaNet, we also use Focal Loss as our classification loss function. We measure our performance on PASCAL VOC2007 and PASCAL VOC2012. Our detector with 300x300 input achieves 73.8% mAP on VOC2007 test, 71.4% mAP on VOC2012 test. And our detector can run at 97FPS and the number of parameters is only 7.7M that meets the requirements of real-time detection. The accuracy of our detector is close to SSD, our detector uses about only 1/3 parameters to SSD.