[OD] DBNet

GitHub - WenmuZhou/DBNet.pytorch: A pytorch re-implementation of Real-time Scene Text Detection with Differentiable Binarization

 

• Our model is based on Cascade R-CNN[2] for Object Detection .
• In detail, we follow the setting get better localization performance through cascading refine boxes.
• Deformable Convolution(DCN)[4] with bottleneck ratio of 4 is applied both on the layer2,layer3 and layer4 of ResNext-101, and
• We use data augmentation(RandomFlip, ShiftScaleRotate, Multi-Scale, CenterCrop) and
• image process algorithm Gaussian noise, RandomBrightnessContrast, Cutout to expand the dataset.
• After analyzing the dataset and prediction results, we mark the easily confused annotation as is crowd and other classes as is crowd.
• In this way, we can only do loss without back propagation, so as not to have a negative impact on the model.
• In addition, we also improved the crop function. If we cut the images directly, it will have a great impact on the boundary, we only keep the box witch IOU greater than 0.8 with original box.