ROI-wise reverse reweighting network for road marking detection

Road markings provide essential navigation information for automated driving and driver assistant systems. As road markings vary dramatically in scale and shape, different markings in various traffic scenes may be sensitive to different level of deep convolutional features. Practically, it is difficult to accurately define which layers are more useful for detecting a marking. It is also inadvisable to extract the same distributed multi-layer features for all ROIs like previous methods (ROI means the region of interest that probably contains markings), which ignores the differences of different markings. To remedy this problem, we propose a novel ROI-wise Reverse Reweighting Network (R³Net) to adaptively combine multi-layer features for different markings. It consists of a multilayer pooling operation and a ROI-wise reverse reweighting module, which independently generates a specific distribution over multi-layer features for each ROI to model different ROI's unique properties. To evaluate our method, we construct a large dataset that contains about 10000 images and 13 categories. Experimental results demonstrate that our proposed network is more effective than others for using multi-layer features.