Robustness of Mask-RCNN at High Ambiguous Environments

Since instance segmentation are popular in the field of unmanned driving, object recognition and instance segmentation in the case of high ambiguity are lacking at the present stage, this paper compares the training results of mask-RCNN algorithm under normal conditions and low visibility with fuzzy conditions. Firstly, for the dataset part, video downloaded from YouTube and be divided into pictures frame by frame by video processing software, which are used as datasets under normal conditions. Then, images are blurred by Fourier transform, which are used as training sets under low visibility conditions. After that, the dataset needs to be normalization, which is to divide the pixel value of the picture by 255 and map them to 0-1 and to improve the training speed. Then we test the data through mask-RCNN, through the tensor Flow library, and train the results on Colabortary. And this algorithm is based on Feature Pyramid Network (FPN) and ResNet101 backbone. Meanwhile there is an Area of Interest (ROI) Pool layer that processes of different sizes into a uniform size output. Through the test of these two datasets, the results show that mask-RCNN's test accuracy can reach more than 99% under normal visibility, and the error rate is extremely low.