Presentation Outline. Semantic Segmentation. Overview. Presentation Outline CNN. Learning Deconvolution Network for Semantic Segmentation 6/6/16

Transcription

1 6/6/16 Learning Deconvolution Network for Semantic Segmentation Hyeonwoo Noh, Seunghoon Hong,Bohyung Han Department of Computer Science and Engineering, POSTECH, Korea Shai Rozenberg 6/6/ Semantic Segmentation Overview Associating each pixel a pre-defined class label Semantic Motion Segmentation U s ing Dens e CRF Formulation,Prateek Singhal, 2014 Training a deconvolution network to perform semantic segmentation. 3 Decoupled Deep N eural N etwork for Semi-s upervis ed Semantic Segmentation, Hyeonwoo N oh, CNN Convolution Subsampling Convolution Subsampling 5 6 1

2 CNN Pooling Layer: non-linear down-sampling layer used to reduce spatial size. CNN Rectified Linear Units Layer: Applies the activation function. f(x)= max(0,x) 7 8 Deconvolution Neural Network Reconstruction an image from the classification vector. Deconvolution Neural Network Reconstruction an image from the classification vector DCNN Unpooling Layer: Reconstruct the original size activation. DCNN Deconvolution Layer: Densify the sparse activations obtained by unpooling

3 Fully Convolutional Network (FCN) Jonathan Long, et al FCN FCN FCN Conditional Random Fields (CRFs) Defining a conditional probability distribution over label sequences, rather than a joint distribution over both label and observation sequences. CRF PXY (, ) Joint disttribution X Observations (Pixels) Y Lables

4 CRF PY ( X) Conditional Distribution PXY (, ) 1 PY ( X) = = Ψc( Xc, Yc) PX ( ) ZX ( ) c C X Observations (Pixels) Y Lables C Cliques Algorithm [Hyeonwoo Noh, et al. 2015] Algorithm Algorithm Several Evolutions: Algorithm Evolutions DeconvNet : EDeconvNet DeconvNet+CRF EDeconvNet+CRF

5 Training Training is done on PASCAL 2012 dataset 2.9M images 250x250 images, 20 classes took 6 days on Nvidia TitenX Training Training is a great challenge as the network s depth leads to significant number of parameters. Batch Normalization: Normalizing each input channel to standard Gaussian distribution. Training Two Stage Training: To improve convergence rate, training would first be done with easy examples and than with challenging examples Goal Results Example of received filters

6 Results Results Conclusions A novel semantic segmentation algorithm by learning a deconvolution network. References [1] Noh, Hyeonwoo, SeunghoonHong, and Bohyung Han. "L earning deconvolution network for semantic segmentation." Proceedings of the IEEE International Conference on Computer Vision [2] Ioffe, Sergey, and Christian Szegedy. "Batch normalization: Accelerating deep network training by reducing internal covariate shift." arxiv preprint arxiv: (2015). Ensemble approach of FCN + CRF. [3] Chen, Liang-Chieh, et al. "Semantic image segmentation with deep convolutional nets and fully connected crfs." arxiv preprint arxiv: (2014). State-of-the-art performance in PASCAL VOC [4] Long, Jonathan, Evan Shelhamer, and Trevor Darrell. "Fully convolutional networks for semantic segmentation." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition [5] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected CRFs. In ICLR,