Dense-SSD for Detecting Small Objects
Chao Li, Youdong Ding, Dan-bei Wang · DEStech Transactions on Engineering and Technology Research · 2018
Object detection with an end-to-end single neural network has achieved great performance both in accuracy and speed. In this paper, we analyze the drawbacks of these methods that they are struggling in small object detection. Next, we propose an improved framework which is based on the state-of-the-art detection architecture, SSD. In contrast to SSD, we replace the original added convolutional layers with dense skip connections to expand the receptive filed of the whole network. This dense connected architecture is proven to be significant to enhance the performance of detecting small objects. The experimental results are shown on PASCAL VOC. Compared to SSD, our dense-SSD has a better detection performance on small objects like aeroplane, bird, bicycle etc. and achieves 78.2 mAP on VOC2007 test with input size of 512. Besides, due to the fewer parameters of the network, our model is smaller than the existing methods with comparable accuracy and speed, nearly about ~70M.