Over the past few years computer vision research has made large strides thanks to the advent of deep learning. Starting from breakthrough results in image classification five years ago, there is currently no computer vision problem that has not been affected by this paradigm shift. Our goal has been to capture a snapshot of the fast-paced research that takes place across the broad spectrum of problems at the interface of deep learning and computer vision. In this context, we are pleased to present a CVIU special issue on Deep Learning for Computer Vision. We have received a total of 34 paper submissions from 19 different countries (Brazil, Canada, China, Czech Republic, Egypt, France, Germany, India, Italy, Japan, Malaysia, Russia, Serbia, Spain, Portugal, Switzerland, Tunisia, UK, USA). The submissions went through an initial check by the guest editors for suitability to the topic, and a few of the submissions were immediately rejected because they were considered o -topic. The remaining papers went through the standard review process with up to three rounds of revisions for some papers. In the end, 12 papers were considered suitable for publication in this special issue. The accepted publications reflect the exciting research that is currently taking place at the interface of deep learning and computer vision. In SMC Faster R-CNN: Toward a Scene-Specialized Multi-Object Detector by Ala Mhalla, Thierry Chateau, Houda Maamatou, Sami Gazzah, Najoua Essoukri and Ben Amaraa, the Sequential Monte Carlo (SCM) tech-nique is used to adapt a generic Faster R-CNN detector to a particular camera environment capturing a traffic scene. When combined with spatio-temporal processing this yields substantial improvements over the generic Faster-RCNN baseline. In Systematic Evaluation of Convolution Neural Network Advances on the ImageNet by Dmytro Mishkin, Jiri Matas and Nikolay Sergievskiy, the authors perform a systematic ablation study of the impact of design choices in deep convolutional neural networks, involving the nonlinearities, pooling functions, learning rates, batch sizes, and normalization options. The results of the paper show that the result of combining the best individual choices yields some-thing only slightly better than the sum of improvements, suggesting that one can explore the impact of every such choices individually. In Speedup of Deep Learning Ensembles for Semantic Segmentation Using a Model Compression Technique by Andrew Holliday, Mo-hammadamin Barekatain, Johannes Laurmaa, Chetak Kandaswamy, Helmut Prendinger, and Helmut Prendinger, the authors apply model compression to the problem of semantic segmentation, demonstrating that compressed models have real-time performance and similar accuracy to ensembles, compression works on ensembles containing entirely distinct deep architectures, while a good upscaling technique matters more than network depth for semantic segmentation. In Compact Descriptors for Sketch-based Image Retrieval using a Triplet loss Convolutional Neural Network by Tu Bui, Leonardo Riberio, Moacir Ponti and John Philip Collomosse, the authors present an efficient representation for sketch based image retrieval derived from a convolutional neural network trained with the triplet loss. The learned representation is shown to generalize to novel categories, while the authors also report stateof-the-art retrieval results while indexing based on compact, 56-bit descriptors. In Deep Compare: A Study on Using Convolutional Neural Net-works to Compare Image Patches by Sergey Zagoruyko and Nikos Ko-modakis, the authors learn a general similarity function for image patches directly from data, exploring Siamese, 2-channel, 2-stream networks and combinations thereof, as well as normalized cross-correlation which are particularly designed for the correspondence task. The authors show that the proposed convolutional networks outperform hand-crafted features by a large margin. In Hand Pose Estimation through Semi-Supervised and Weakly-Supervised Learning by Natalia Neverova, Christian Wolf, Florian Nebout and Graham Taylor, the authors introduce a hand pose estimation method that relies on direct regression from depth and segmentation maps. Building on this, the authors develop an intermediate representation that allows the adaptation from synthetic to real data, while also developing a restorationdriven method that yields a supervision signal on unlabeled data. In Weak/semi supervised learning: Harnessing Noisy Web Images for Deep Representation by Phong Dinh Vo, Alexandru Ginsca, Herve Le Borgne and Adrian Popescu, the authors pursue semi-supervised learning approaches to exploit large amounts of unannotated images downloaded from Flickr and Bing. The authors show that one can learn useful representations de-spite the high level of noise in training data, and demonstrate the transferability of the learned representations to new problems. In Saliency Driven Object Recognition in Egocentric Videos with Deep CNN: toward application in assistance to Neuroprostheses by Philippe Perez de San Roman, Jenny Benois-Pineau, Jean-Philippe Domenger, Florent Paclet, Daniel Cattaert and Aymar de Rugy, the authors propose an