Object Detection For Autonomous vehicles(CitySegNet)

We use DUC (Dense upsampling convulution ) which is a CNN based model for semantic segmentation which uses an image classification network (ResNet) as a backend and achieves improved accuracy in terms of mIOU score using two novel techniques. The first technique is called Dense Upsampling Convolution (DUC) which generates pixel-level prediction by capturing and decoding more detailed information that is generally missing in bilinear upsampling. Secondly, a framework called Hybrid Dilated Convolution (HDC) is proposed in the encoding phase which enlarges the receptive fields of the network to aggregate global information. It also alleviates the checkerboard receptive field problem (“gridding”) caused by the standard dilated convolution operation.

Website