![]() ![]() The top-down pathway and lateral connections are visualized in Fig 3. This process is iterated until each feature map from the bottom-up pathway has a corresponding new feature map connected with lateral connections. These two feature maps are then merged 3 by element-wise addition to form a new feature map. Using nearest neighbor upsampling, the last feature map from the bottom-up pathway is expanded to the same scale as the second-to-last feature map. Top-down pathway and lateral connections. The bottom-up pathway is visualized in Fig 2. These chosen feature maps will be used as the foundation of the feature pyramid. The bottom-up pathway of building FPN is accomplished by choosing the last feature map of each group of consecutive layers 2 that output feature maps of the same scale. Recall that in ResNet, some consecutive layers may output feature maps of the same scale but generally, feature maps of deeper layers have smaller scales/resolutions. They are described as below.īottom-up pathway. The construction of FPN involves two pathways which are connected with lateral connections. The fully convolutional nature enables the network to take an image of an arbitrary size and outputs proportionally sized feature maps at multiple levels in the feature pyramid. (2017) as its backbone, which is in turn built on top of ResNet (ResNet-50, ResNet-101 or ResNet-152) 1 in a fully convolutional fashion. RetinaNet adopts the Feature Pyramid Network (FPN) proposed by Lin, Dollar, et al. Note: for a brief introduction and comparison among popular detectors before RetinaNet (e.g., R-CNN), see (Tryolabs 2017 Xu 2017) I have also found a post by Hollemans (2018) to be very informative. Therefore, I read the original paper and many related ones carefully and post shares what I have learnt. To use the detector appropriately, I need to study its design and intuitions. Therefore, RetinaNet appears to be an ideal candidate for the project. According to the paper, RetinaNet showed both ideal accuracy and speed compared to other detectors while still keeping a very simple construct plus, there is an opensource implementaion by Gaiser et al. 2013), Fast R-CNN (Girshick 2015), SSD (Liu et al. I found several popular detectors including: OverFeat (Sermanet et al. Recently I have been doing some research on object detection, trying to find a state-of-the-art detector for a project. ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |