Object Detection Based On Deformable Convolutional Neural Networks

In recent years,convolution neutral network(CNN)and Deformable Part-based Model(DPM)have been successfully applied to computer visual recognition task.As a type of deep neural network,CNN combines multilayer neural network and convolution computing,and through sparse connection,weights sharing and pooling,they greatly reduce the model parameters and make it suitable for large scale computer learning and training task while achieving high robustness and excellent recognition.Based on graphical structure model,DPM inherent the concept of “decompose object as parts” and learns parts filter and model parameters from weak tagged datasets by introducing latent support vector machine(latent SVM).In order to reflect variance within cluster,it brings in mixed components,and realized high generalization and finally achieved 2010 PASCAL VOC life achievement reward.This paper focus on the relation between CNN and DPM,and proposed a novel structural model that integrates a deformable layer and CNN model.The parts filter coupling with deformable layer,generalizes CNN to multi-parts structure,eventually make the model more robust with in cluster variation.In contrast to previous approach,the part filters which we called “multi-parts filter” are discovered from the CNN high layer feature descriptor layer that already contains parts information.Each part is consisted of multiple filters which can fully describe part information from different perspectives.In order to find the part filters,gradient maps are applied for localizing of parts and analyze the gradient maps of the network outputs and find relation between the activation centers and annotated semantic parts and bounding boxes.We also conducted part localizing experiments on CUB200-2011 dataset.The experimental result verified the feasibility of the method.Base on “multi-part filter” model and CNN that introduced deformable layer,deformable convolutional neural network model(DCNN)is proposed,which can improve the robustness to deformable objects.Finally,we tested our model on PASCAL VOC 2007 datasets,and the result shows our deformable CNN model can efficiently increase the detection success rate.