ECCV 2018 Papers

Although deep learning approaches have stood out in recent years due to their state-of-the-art results, they continue to suffer from catastrophic forgetting, a dramatic decrease in overall performance when training with new classes added incrementally. This is due to current neural network architectures requiring the entire dataset, consisting of all the samples from the old as well as the new classes, to update the model -a requirement that becomes easily unsustainable as the number of classes grows. We address this issue with our approach to learn deep neural networks incrementally, using new data and only a small exemplar set corresponding to samples from the old classes. This is based on a loss composed of a distillation measure to retain the knowledge acquired from the old classes, and a cross-entropy loss to learn the new classes. Our incremental training is achieved while keeping the entire framework end-to-end, i.e., learning the data representation and the classifier jointly, unlike recent methods with no such guarantees. We evaluate our method extensively on the CIFAR-100 and ImageNet (ILSVRC 2012) image classification datasets, and show state-of-the-art performance.

In this paper we address the problem of detecting crosswalks from LiDAR and camera imagery. Towards this goal, given multiple LiDAR sweeps and the corresponding imagery, we project both inputs onto the ground surface to produce a top down view of the scene. We then leverage convolutional neural networks to extract semantic cues about the location of the crosswalks. These are then used in combination with road centerlines from freely available maps (e.g., OpenStreetMaps) to solve a structured optimization problem which draws the final crosswalk boundaries. Our experiments over crosswalks in a large city area show that 96.6% automation can be achieved.

Spatial pyramid pooling module or encode-decoder structure are used in deep neural networks for semantic segmentation task. The former networks are able to encode multi-scale contextual information by probing the incoming features with filters or pooling operations at multiple rates and multiple effective fields-of-view, while the latter networks can capture sharper object boundaries by gradually recovering the spatial information. In this work, we propose to combine the advantages from both methods. Specifically, our proposed model, DeepLabv3+, extends DeepLabv3 by adding a simple yet effective decoder module to refine the segmentation results especially along object boundaries. We further explore the Xception model and apply the depthwise separable convolution to both Atrous Spatial Pyramid Pooling and decoder modules, resulting in a faster and stronger encoder-decoder network. We demonstrate the effectiveness of the proposed model on PASCAL VOC 2012 and Cityscapes datasets, achieving the test set performance of 89% and 82.1% without any post-processing. Our paper is accompanied with a publicly available reference implementation of the proposed models in Tensorflow at url{https://github.com/tensorflow/models/tree/master/research/deeplab}.

Recent automotive vision work has focused almost exclusively on processing forward-facing cameras. However, future autonomous vehicles will not be viable without a more comprehensive surround sensing, akin to a human driver, as can be provided by 360° panoramic cameras. We present an approach to adapt contemporary deep network architectures developed on conventional rectilinear imagery to work on equirectangular 360° panoramic imagery. To address the lack of annotated panoramic automotive datasets availability, we adapt contemporary automotive dataset, via style and projection transformations, to facilitate the cross-domain retraining of contemporary algorithms for panoramic imagery. Following this approach we retrain and adapt existing architectures to recover scene depth and 3D pose of vehicles from monocular panoramic imagery without any panoramic training labels or calibration parameters. Our approach is evaluated qualitatively on crowd-sourced panoramic images and quantitatively using an automotive environment simulator to provide the first benchmark for such techniques within panoramic imagery.

Many classical Computer Vision problems, such as essential matrix computation and pose estimation from 3D to 2D correspondences, can be solved by finding the eigenvector corresponding to the smallest, or zero, eigenvalue of a matrix representing a linear system. Incorporating this in deep learning frameworks would allow us to explicitly encode known notions of geometry, instead of having the network implicitly learn them from data. However, performing eigendecomposition within a network requires the ability to differentiate this operation. Unfortunately, while theoretically doable, this introduces numerical instability in the optimization process in practice. In this paper, we introduce an eigendecomposition-free approach to training a deep network whose loss depends on the eigenvector corresponding to a zero eigenvalue of a matrix predicted by the network. We demonstrate on several tasks, including keypoint matching and 3D pose estimation, that our approach is much more robust than explicit differentiation of the eigendecomposition, It has better convergence properties and yields state-of-the-art results on both tasks.

Predicting future activities from an egocentric viewpoint is of particular interest in assisted living. However, state-of-the-art egocentric activity understanding techniques are mostly NOT capable of predictive tasks, as their synchronous processing architecture performs poorly in either modeling event dependency or pruning temporal redundant features. This work explicitly addresses these issues by proposing an asynchronous gaze-event driven attentive activity prediction network. This network is built on a gaze-event extraction module inspired by the fact that gaze moving in/out a certain object most probably indicates the occurrence/ending of a certain activity. The extracted gaze events are input to: 1) an asynchronous module which reasons about the temporal dependency between events and 2) a synchronous module which softly attends to informative temporal durations for more compact and discriminative feature extraction. Both modules are seamlessly integrated for collaborative prediction. Extensive experimental results on egocentric activity prediction as well as recognition well demonstrate the effectiveness of the proposed method.

Uncertainty estimation in deep learning becomes more important recently. A deep learning model can't be applied in real applications if we don't know whether the model is certain about the decision or not. Some literature proposes the Bayesian neural network which can estimate the uncertainty by Monte Carlo Dropout (MC dropout). However, MC dropout needs to forward the model N times which results in N times slower. For real-time applications such as a self-driving car system, which needs to obtain the prediction and the uncertainty as fast as possible, so that MC dropout becomes impractical. In this work, we propose the region-based temporal aggregation (RTA) method which leverages the temporal information in videos to simulate the sampling procedure. Our RTA method with Tiramisu backbone is 10x faster than the MC dropout with Tiramisu backbone (N=5). Furthermore, the uncertainty estimation obtained by our RTA method is comparable to MC dropout's uncertainty estimation on pixel-level and frame-level metrics.

Template matching is a fundamental problem in computer vision, with many applications. Existing methods use sliding window computation for choosing an image-window that best matches the tem- plate. For classic algorithms based on SSD, SAD and normalized cross- correlation, efficient algorithms have been developed allowing them to run in real-time. Current state of the art algorithms are based on nearest neighbor (NN) matching of small patches within the template to patches in the image. These algorithms yield state-of-the-art results since they can deal better with changes in appearance, viewpoint, illumination, non- rigid transformations, and occlusion. However, NN-based algorithms are relatively slow not only due to NN computation for each image patch, but also since their sliding window computation is inecient. We there- fore propose in this paper an efficient NN-based algorithm. Its accuracy is similar (in some cases slightly better) than the existing algorithms and its running time is 44-200 times faster depending on the sizes of the images and templates used. The main contribution of our method is an algorithm for incrementally computing the score of each image window based on the score computed for the previous window. This is in con- trast to computing the score for each image window independently, as in previous NN-based methods. The complexity of our method is there- fore O(|I|) instead of O(|I||T|), where I and T are the image and the template respectively.

We introduce Spatial Group Convolution (SGC) for accelerating the computation of 3D dense prediction tasks. SGC is orthogonal to group convolution, which works on spatial dimensions rather than feature channel dimension. It divides input voxels into different groups, then conducts 3D sparse convolution on these separated groups. As only valid voxels are considered when performing convolution, computation can be significantly reduced with a slight loss of accuracy. The proposed operations are validated on semantic scene completion task, which aims to predict a complete 3D volume with semantic labels from a single depth image. With SGC, we further present an efficient 3D sparse convolutional network, which harnesses a multiscale architecture and a coarse-to-fine prediction strategy. Evaluations are conducted on the SUNCG dataset, achieving state-of-the-art performance and fast speed.

A sizable body of work on relative attributes provides compelling evidence that relating pairs of images along a continuum of strength pertaining to a visual attribute yields significant improvements in a wide variety of tasks in vision. In this paper, we show how emerging ideas in graph neural networks can yield a unified solution to various problems that broadly fall under relative attribute learning. Our main idea is the realization that relative attribute learning naturally benefits from exploiting the graphical structure of dependencies among the different relative attributes of images, especially when only partial ordering of the relative attributes is provided in the training data. We use message passing on a probabilistic graphical model to perform end to end learning of appropriate representations of the images, their relationships as well as the interplay between different attributes to best align with provided annotations. Our experiments demonstrate that this simple end-to-end learning framework using GNNs is very effective in achieving competitive accuracy with specialized methods for both relative attribute learning and binary attribute prediction, while significantly relaxing the requirements on the training data and/or the number of parameters or both.

Three-dimensional rigid point cloud registration has many applications in computer vision and robotics. Local methods tend to fail, causing global methods to be needed, when the relative transformation is large or the overlap ratio is small. Most existing global methods utilize BnB optimization over the 6D parameter space of SE(3). Such methods are usually very slow because the time complexity of BnB optimization is exponential in the dimensionality of the parameter space. In this paper, we decouple the optimization of translation and rotation, and we propose a fast BnB algorithm to globally optimize the 3D translation parameter first. The optimal rotation is then calculated by utilizing the global optimal translation found by the BnB algorithm. The separate optimization of translation and rotation is realized by using a newly proposed rotation invariant feature. Experiments on challenging data sets demonstrate that the proposed method outperforms state-of-the-art global methods in terms of both speed and accuracy.

Most existing CNN-based methods for single-view 3D object reconstruction represent a 3D object as either a 3D voxel occupancy grid or multiple depth-mask image pairs. However, these representations are inefficient since empty voxels or background pixels are wasteful. We propose a novel approach that addresses this limitation by replacing masks with ''deformation-fields''. Given a single image at an arbitrary viewpoint, a CNN predicts multiple surfaces, each in a canonical location relative to the object. Each surface comprises a depth-map and corresponding deformation-field that ensures every pixel-depth pair in the depth-map lies on the object surface. These surfaces are then fused to form the full 3D shape. During training, we use a combination of per-view and multi-view losses. The novel multi-view loss encourages the 3D points back-projected from a particular view to be consistent across views. Extensive experiments demonstrate the efficiency and efficacy of our method on single-view 3D object reconstruction.

Virtual and augmented reality technologies have seen significant growth in the past few years. A key component of such systems is the ability to track the pose of head mounted displays and controllers in 3D space. We tackle the problem of efficient 6-DoF tracking of a handheld controller from egocentric camera perspectives. We collected the HMD Controller dataset which consist of over 540,000 stereo image pairs labelled with the full 6-DoF pose of the handheld controller. Our proposed SSD-AF-Stereo3D model achieves a mean average error of 33.5 millimeters in 3D keypoint prediction and is used in conjunction with an IMU sensor on the controller to enable 6-DoF tracking. We also present results on approaches for model based full 6-DoF tracking. All our models operate under the strict constraints of real time mobile CPU inference.

Training a deep network to perform semantic segmentation requires large amounts of labeled data. To alleviate the manual effort of annotating real images, researchers have investigated the use of synthetic data, which can be labeled automatically. Unfortunately, a network trained on synthetic data performs relatively poorly on real images. While this can be addressed by domain adaptation, existing methods all require having access to real images during training. In this paper, we introduce a drastically different way to handle synthetic images that does not require seeing any real images at training time. Our approach builds on the observation that foreground and background classes are not affected in the same manner by the domain shift, and thus should be treated differently. In particular, the former should be handled in a detection-based manner to better account for the fact that, while their texture in synthetic images is not photo-realistic, their shape looks natural. Our experiments evidence the effectiveness of our approach on Cityscapes and CamVid with models trained on synthetic data only.

We introduce a fast and efficient convolutional neural network, ESPNet, for semantic segmentation of high resolution images under resource constraints. ESPNet is based on a new convolutional module, efficient spatial pyramid (ESP), which is efficient in terms of computation, memory, and power. ESPNet is 22 times faster (on a standard GPU) and 180 times smaller than the state-of-the-art semantic segmentation network PSPNet, while its category-wise accuracy is only 8% less. We evaluated EPSNet on a variety of semantic segmentation datasets including Cityscapes, PASCAL VOC, and a breast biopsy whole slide image dataset. Under the same constraints on memory and computation, ESPNet outperforms all the current efficient CNN networks such as MobileNet, ShuffleNet, and ENet on both standard metrics and our newly introduced performance metrics that measure efficiency on edge devices. Our network can process high resolution images at a rate of 112 and 9 frames per second on a standard GPU and edge device, respectively.

Recent studies on face attribute transfer have achieved great success. A lot of models are able to transfer face attributes with an input image. However, they suffer from three limitations: (1) incapability of generating image by exemplars; (2) being unable to transfer multiple face attributes simultaneously; (3) low quality of generated images, such as low-resolution or artifacts. To address these limitations, we propose a novel model which receives two images of opposite attributes as inputs. Our model can transfer exactly the same type of attributes from one image to another by exchanging certain part of their encodings. All the attributes are encoded in a disentangled manner in the latent space, which enables us to manipulate several attributes simultaneously. Besides, our model learns the residual images so as to facilitate training on higher resolution images. With the help of multi-scale discriminators for adversarial training, it can even generate high-quality images with finer details and less artifacts. We demonstrate the effectiveness of our model on overcoming the above three limitations by comparing with other methods on the CelebA face database. A pytorch implementation is available at https://github.com/Prinsphield/ELEGANT.

The state of the art in video understanding suffers from two problems: (1) The major part of reasoning is performed locally in the video, thus missing important relationships within actions that span several seconds. (2) While there are local methods with fast per-frame processing, the processing of the whole video is not efficient and hampers fast video retrieval or online classification of long-term activities. In this paper, we introduce a network architecture that takes long-term content into account and enables fast per-video processing at the same time. The architecture is based on merging long-term content already in the network rather than in a post-hoc fusion. Together with a sampling strategy, which exploits that neighboring frames are largely redundant, this yields high-quality action classification and video captioning at up to 230 videos per second, where each video can consist of a few hundred frames. The approach achieves competitive performance across all datasets while being 10x to 80x faster than state-of-the-art methods.

Point clouds obtained from 3D scans are typically sparse, irregular, and noisy, and required to be consolidated. In this paper, we present the first deep learning based {em edge-aware} technique to facilitate the consolidation of point clouds. We design our network to process points grouped in local patches, and train it to learn and help consolidate points, deliberately for edges. To achieve this, we formulate a regression component to simultaneously recover 3D point coordinates and point-to-edge distances from upsampled features, and an edge-aware joint loss function to directly minimize distances from output points to 3D meshes and to edges. Compared with previous neural network based works, our consolidation is {em edge-aware/}. During the synthesis, our network can attend to the detected sharp edges and enable more accurate 3D reconstructions. Also, we trained our network on virtual scanned point clouds, demonstrated the performance of our method on both synthetic and real point clouds, presented various surface reconstruction results, and showed how our method outperforms the state-of-the-arts.

We propose dynamic task prioritization for multitask learning. This allows a model to dynamically prioritize difficult tasks during training, where difficulty is inversely proportional to performance, and where difficulty changes over time. In contrast to curriculum learning, where easy tasks are prioritized above difficult tasks, we present several studies showing the importance of prioritizing difficult tasks first. We observe that imbalances in task difficulty can lead to unnecessary emphasis on easier tasks, thus neglecting and slowing progress on difficult tasks. Motivated by this finding, we introduce a notion of dynamic task prioritization to automatically prioritize more difficult tasks by adaptively adjusting the mixing weight of each task's loss objective. Additional ablation studies show the impact of the task hierarchy, or the task ordering, when explicitly encoded in the network architecture. Our method outperforms existing multitask methods and demonstrates competitive results with modern single-task models on the COCO and MPII datasets.