Abstract

No Results Found

Abstract

Efficient CNN designs like ResNets and DenseNet were proposed to improve accuracy vs efficiency trade-offs. They essentially increased the connectivity, allowing efficient information flow across layers. Inspired by these techniques, we propose to model connections between filters of a CNN using graphs which are simultaneously sparse and well connected. Sparsity results in efficiency while well connectedness can preserve the expressive power of the CNNs. We use a well-studied class of graphs from theoretical computer science that satisfies these properties known as Expander graphs. Expander graphs are used to model connections between filters in CNNs to design networks called X-Nets. We present two guarantees on the connectivity of X-Nets: Each node influences every node in a layer in logarithmic steps, and the number of paths between two sets of nodes is proportional to the product of their sizes. We also propose efficient training and inference algorithms, making it possible to train deeper and wider X-Nets effectively. Expander based models give a 4% improvement in accuracy on MobileNet over grouped convolutions, a popular technique, which has the same sparsity but worse connectivity. X-Nets give better performance trade-offs than the original ResNet and DenseNet-BC architectures. We achieve model sizes comparable to state-of-the-art pruning techniques using our simple architecture design, without any pruning. We hope that this work motivates other approaches to utilize results from graph theory to develop efficient network architectures.

Abstract

Deep neural networks are highly effective at a range of computational tasks. However, they tend to be computationally expensive, especially in vision-related problems, and also have large memory requirements. One of the most effective methods to achieve significant improvements in computational/spatial efficiency is to binarize the weights and activations in a network. However, naive binarization results in accuracy drops when applied to networks for most tasks. In this work, we present a highly generalized, distribution-aware approach to binarizing deep networks that allows us to retain the advantages of a binarized network, while reducing accuracy drops. We also develop efficient implementations for our proposed approach across different architectures. We present a theoretical analysis of the technique to show the effective representational power of the resulting layers, and explore the forms of data they …

Abstract

Binarization is an extreme network compression approach that provides large computational speedups along with energy and memory savings, albeit at significant accuracy costs. We investigate the question of where to binarize inputs at layer-level granularity and show that selectively binarizing the inputs to specific layers in the network could lead to significant improvements in accuracy while preserving most of the advantages of binarization. We analyze the binarization tradeoff using a metric that jointly models the input binarization-error and computational cost and introduce an efficient algorithm to select layers whose inputs are to be binarized. Practical guidelines based on insights obtained from applying the algorithm to a variety of models are discussed. Experiments on Imagenet dataset using AlexNet and ResNet-18 models show 3-4% improvements in accuracy over fully binarized networks with minimal

Abstract

We, humans, have the ability to easily imagine scenes that depict sentences such as “Today is a beautiful sunny day” or “There is a Christmas feel, in the air”. While it is hard to precisely describe what one person may imagine, the essential high-level themes associated with such sentences largely remains the same. The ability to synthesize novel images that depict the feel of a sentence is very useful in a variety of applications such as education, advertisement, and entertainment. While existing papers tackle this problem given a style image, we aim to provide a far more intuitive and easy to use solution that synthesizes novel renditions of an existing image, conditioned on a given sentence. We present a method for cross-modal style transfer between an English sentence and an image, to produce a new image that imbibes the essential theme of the sentence. We do this by modifying the style transfer mechanism

Abstract

A system for capturing horizontal disparity stereo panorama is disclosed . The system includes a multi surface selective light reflector unit , a secondary reflector and a computing unit . The multi surface selective light reflector unit ( a ) obtains light rays from a 3D scene of outside world that are relevant to create ( i ) a left eye panorama and ( ii ) a right eye panorama and ( b ) reflects the light rays without internal reflections between the light rays . The secondary reflector ( a ) obtains the reflected light rays from the multi surface selective light reflector unit and ( b ) reflects the light rays through the viewing aperture . The computing unit captures ( i ) the reflected light rays from the secondary reflector and ( ii ) the upper part of the 3D scene from a concave lens as a warped image and processes the warped image to ( a ) the left eye panorama and ( b ) the right eye panorama .

Abstract

We look at the problem of developing a compact and accurate model for gesture recognition from videos in a deep-learning framework. Towards this we propose a joint 3DCNN-LSTM model that is end-to-end trainable and is shown to be better suited to capture the dynamic information in actions. The solution achieves close to state-of-the-art accuracy on the ChaLearn dataset, with only half the model size. We also explore ways to derive a much more compact representation in a knowledge distillation framework followed by model compression. The final model is less than 1 MB in size, which is less than one hundredth of our initial model, with a drop of 7% in accuracy, and is suitable for real-time gesture recognition on mobile devices.

Abstract

wo-view stereo problem is a well researchedproblem in 3D computer vision. Algorithms proposedin the past have focussed on rectified stereo imageswhere the epipolar lines are parallel to the horizon-tal axis.The general problem of computing stereocorrespondences for unrectified images without anyknowledge of calibration parameters is an importantproblem but unexplored as yet. Our idea in this paperis to predict depth maps from two unrectified stereoimages using a modified Flownet architecture. Since,datasets for depth map reconstruction for unrectifiedstereo images for deep learning do not exist, we havecreated a dataset of turn table sequences of 3D modelsfrom Google 3D warehouse. Following the conceptsof Attention modelling, we implement an architecturefor combining correlations computed at multipleresolutions using a simple element-wise multiplicationof the correlations to aid the architecture to resolvecorrespondences for textureless and repeated texturedsurfaces. Our experiments show both qualitaitve andquantitative improvements of depth maps over theoriginal Flownet architecture.

Abstract

Video highlights are a selection of the most interesting parts of a video. The problem of highlight de-tection has been explored for video domains like egocentric,sports, movies, and surveillance videos. Existing methods are limited to finding visually important parts of the video but does not necessarily learn semantics. Moreover, the available benchmark datasets contain audio muted, single activity,short videos, which lack any context apart from a few keyframes that can be used to understand them. In this work,we explore highlight detection in the TV series domain, which features complex interactions with the surroundings. The existing methods would fare poorly in capturing the video semantics in such videos. To incorporate the importance of dialogues/audio, we propose using the descriptions of shots of the video as cues to learning visual importance.Note that while the audio information is used to determine visual importance during training, the highlight detection still works using only the visual information from videos.We use publicly available text ranking algorithms to rank the descriptions. The ranking scores are used to train a visual pair wise shot ranking model (VPSR) to find the highlights of the video. The results are reported on TV series videos of the Video Set dataset and a season of Buffy the Vampire SlayerTV series.

Abstract

Person recognition methods that use multiple body regions have shown significant improvements over traditional face-based recognition. One of the primary challenges in full-body person recognition is the extreme variation in pose and view point. In this work,(i) we present an approach that tackles pose variations utilizing multiple models that are trained on specific poses, and combined using pose-aware weights during testing.(ii) For learning a person representation, we propose a network that jointly optimizes a single loss over multiple body regions.(iii) Finally, we introduce new benchmarks to evaluate person recognition in diverse scenarios and show significant improvements over previously proposed approaches on all the benchmarks including the photo album setting of PIPA.