Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers—8× deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation.

/pdf/deep-residual-learning-for-image-recognition-b75jg61qrq.pdf

Deep Residual Learning for Image Recognition

In this work we investigate the effect of the convolutional network depth on its accuracy in the large-scale image recognition setting. Our main contribution is a thorough evaluation of networks of increasing depth using an architecture with very small (3x3) convolution filters, which shows that a significant improvement on the prior-art configurations can be achieved by pushing the depth to 16-19 weight layers. These findings were the basis of our ImageNet Challenge 2014 submission, where our team secured the first and the second places in the localisation and classification tracks respectively. We also show that our representations generalise well to other datasets, where they achieve state-of-the-art results. We have made our two best-performing ConvNet models publicly available to facilitate further research on the use of deep visual representations in computer vision.

Very Deep Convolutional Networks for Large-Scale Image Recognition

Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers---8x deeper than VGG nets but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. 
The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation.

We propose a deep convolutional neural network architecture codenamed Inception that achieves the new state of the art for classification and detection in the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC14). The main hallmark of this architecture is the improved utilization of the computing resources inside the network. By a carefully crafted design, we increased the depth and width of the network while keeping the computational budget constant. To optimize quality, the architectural decisions were based on the Hebbian principle and the intuition of multi-scale processing. One particular incarnation used in our submission for ILSVRC14 is called GoogLeNet, a 22 layers deep network, the quality of which is assessed in the context of classification and detection.

/pdf/going-deeper-with-convolutions-1yobw2o2ds.pdf

Going deeper with convolutions

We present a method for human action recognition from video, which exploits both form (local shape) and motion (local flow). Inspired by models of the human visual system, the two feature sets are processed independently in separate channels. The form channel extracts a dense local shape representation from every frame, while the motion channel extracts dense optic flow from the frame and its immediate predecessor. The same processing pipeline is applied in both channels: feature maps are pooled locally, down-sampled, and compared to a collection of learnt templates, yielding a vector of similarity scores. In a final step, the two score vectors are merged, and recognition is performed with a discriminative classifier. In an evaluation on two standard datasets our method outperforms the state-of-the-art, confirming that the combination of form and motion improves recognition.

Combining Densely Sampled Form and Motion for Human Action Recognition

Comparing shapes is important for both recognition and geometry-based grouping. In the case of perspective views of planar shapes, such comparison will typically be based on projective invariants. Yet there are quite a number of practical cases where one can do better and exploit simpler invariants of subgroups of the planar projectivities. This paper tries to sketch a systematic approach, based on subgroups defined by the structures they keep fixed in the image. In particular, this paper focuses on fixed points, fixed lines, lines of fixed points, and pencils of fixed lines. A complete classification of the resulting subgroups is given and the most interesting cases are identified. For these cases invariants are given and examples illustrate their use. In order to illustrate the wider scope, perspectively skewed mirror symmetry is discussed, since it entails a different type of fixed structures.© (1995) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Groups for grouping

In building 3D site models for visualization and virtual walkthrough, most emphasis so far has been on creating the 3D shape models. Less emphasis has been on creating realistic textures, e.g. to simulate building materials or vegetation. Nevertheless, the appearance of object and landscape models will depend at least as much on their textures, as on the precision of their geometry. The paper proposes a texture synthesis technique for the simulation of building materials and vegetation types. As textures can often more easily be described as a composition of subtextures than as a single texture, a hierarchical extension of our basic texture synthesis is proposed. These "composite textures" are based on a kind of metatexture concept, i.e. the layout of the different subtextures is itself modeled as a texture, that can be generated automatically. Examples are shown for building materials with an intricate structure and for the automatic creation of landscape textures.

Composite textures: emulating building materials and vegetation for 3D models

Two tourist guides are described One supports a virtual tour through an archaeological site, the other a tour through a real exhibition The first system is based on the 3D reconstruction of the ancient city of Sagalassos A virtual guide, represented by an animated mask can be given commands using natural speech Through its expressions the mask makes clear whether the questions have been understood, whether they make sense, etc Its presence largely increases the intuitiveness of the interface This system is described only very concisely

A second system is a palmtop assistant that gives information about the paintings at the ongoing Van Dijck exhibition in the Antwerp museum of fine arts The system consists of a handheld PC with camera and Ethernet radio link Images are taken of paintings or details thereof The images are analysed by a server, which sends back information about the particular painting or the details It gives visitors more autonomy in deciding in which order to look at pieces and how much information is required about each The system is based on image database retrieval, where interest points are characterised by geometric/photometric invariants of their neighbourhoods

/pdf/adventurous-tourism-for-couch-potatoes-x33ax2cyf7.pdf

Adventurous Tourism for Couch Potatoes

Textures can often more easily be described as a composition of subtextures than as a single texture. The paper proposes a way to model and synthesize such "composite textures", where the layout of the different subtextures is itself modeled as a texture, which can be generated automatically. Examples are shown for building materials with an intricate structure and for the automatic creation of landscape textures. First, a model of the composite texture is generated. This procedure comprises manual or unsupervised texture segmentation to learn the spatial layout of the composite texture and the extraction of models for each of the subtextures. Synthesis of a composite texture includes the generation of a layout texture, which is subsequently filled in with the appropriate subtextures. This scheme is refined further by also including interactions between neighboring subtextures.

Luc Van Gool

Papers

Combining Densely Sampled Form and Motion for Human Action Recognition

Groups for grouping

Composite textures: emulating building materials and vegetation for 3D models

Adventurous Tourism for Couch Potatoes

Composite texture descriptions