In generative modeling, the Wasserstein distance (WD) has emerged as a useful metric to measure the discrepancy between generated and real data distributions. Unfortunately, it is challenging to approximate the WD of high-dimensional distributions. In contrast, the sliced Wasserstein distance (SWD) factorizes high-dimensional distributions into their multiple one-dimensional marginal distributions and is thus easier to approximate. In this paper, we introduce novel approximations of the primal and dual SWD. Instead of using a large number of random projections, as it is done by conventional SWD approximation methods, we propose to approximate SWDs with a small number of parameterized orthogonal projections in an end-to-end deep learning fashion. As concrete applications of our SWD approximations, we design two types of differentiable SWD blocks to equip modern generative frameworks---Auto-Encoders (AE) and Generative Adversarial Networks (GAN). In the experiments, we not only show the superiority of the proposed generative models on standard image synthesis benchmarks, but also demonstrate the state-of-the-art performance on challenging high resolution image and video generation in an unsupervised manner.
In the context of deep learning, neural networks with multiple branches have been used that each solve different tasks. Such ramified networks typically start with a number of shared layers, after which different tasks branch out into their own sequence of layers. As the number of possible network configurations is combinatorially large, prior work has often relied on ad hoc methods to determine the level of layer sharing. This work proposes a novel method to assess the relatedness of tasks in a principled way. We base the relatedness of a task pair on the usefulness of a set of features of one task for the other, and vice versa. The resulting task affinities are used for the automated construction of a branched multi-task network in which deeper layers gradually grow more task-specific. Our multi-task network outperforms the state-of-the-art on CelebA. Additionally, the layer sharing schemes devised by our method outperform common multi-task learning models which were constructed ad hoc. We include additional experiments on Cityscapes and SUN RGB-D to illustrate the wide applicability of our approach. Code and trained models for this paper are made available https://github.com/SimonVandenhende/
Learning descriptive spatio-temporal object models from data is paramount for the task of semi-supervised video object segmentation. Most existing approaches mainly rely on models that estimate the segmentation mask based on a reference mask at the first frame (aided sometimes by optical flow or the previous mask). These models, however, are prone to fail under rapid appearance changes or occlusions due to their limitations in modelling the temporal component. On the other hand, very recently, other approaches learned long-term features using a convolutional LSTM to leverage the information from all previous video frames. Even though these models achieve better temporal representations, they still have to be fine-tuned for every new video sequence. In this paper, we present an intermediate solution and devise a novel GAN architecture, FaSTGAN, to learn spatio-temporal object models over finite temporal windows. To achieve this, we concentrate all the heavy computational load to the training phase with two critics that enforce spatial and temporal mask consistency over the last K frames. Then at test time, we only use a relatively light regressor, which reduces the inference time considerably. As a result, our approach combines a high resiliency to sudden geometric and photometric object changes with efficiency at test time (no need for fine-tuning nor post-processing). We demonstrate that the accuracy of our method is on par with state-of-the-art techniques on the challenging YouTube-VOS and DAVIS datasets, while running at 32 fps, about 4x faster than the closest competitor.
Autonomous vehicles are more likely to be accepted if they drive accurately, comfortably, but also similar to how human drivers would. This is especially true when autonomous and human-driven vehicles need to share the same road. The main research focus thus far, however, is still on improving driving accuracy only. This paper formalizes the three concerns with the aim of accurate, comfortable and human-like driving. Three contributions are made in this paper. First, numerical map data from HERE Technologies are employed for more accurate driving; a set of map features which are believed to be relevant to driving are engineered to navigate better. Second, the learning procedure is improved from a pointwise prediction to a sequence-based prediction and passengers' comfort measures are embedded into the learning algorithm. Finally, we take advantage of the advances in adversary learning to learn human-like driving; specifically, the standard L1 or L2 loss is augmented by an adversary loss which is based on a discriminator trained to distinguish between human driving and machine driving. Our model is trained and evaluated on the Drive360 dataset, which features 60 hours and 3000 km of real-world driving data. Extensive experiments show that our driving model is more accurate, more comfortable and behaves more like a human driver than previous methods. The resources of this work will be released on the project page.
We propose a Three-Player Generative Adversarial Network to improve classification networks. In addition to the game played between the discriminator and generator, a competition is introduced between the generator and the classifier. The generator's objective is to synthesize samples that are both realistic and hard to label for the classifier. Even though we make no assumptions on the type of augmentations to learn, we find that the model is able to synthesize realistically looking examples that are hard for the classification model. Furthermore, the classifier becomes more robust when trained on these difficult samples. The method is evaluated on a public dataset for traffic sign recognition.
This work proposes a new method to accurately complete sparse LiDAR maps guided by RGB images. For autonomous vehicles and robotics the use of LiDAR is indispensable in order to achieve precise depth predictions. A multitude of applications depend on the awareness of their surroundings, and use depth cues to reason and react accordingly. On the one hand, monocular depth prediction methods fail to generate absolute and precise depth maps. On the other hand, stereoscopic approaches are still significantly outperformed by LiDAR based approaches. The goal of the depth completion task is to generate dense depth predictions from sparse and irregular point clouds which are mapped to a 2D plane. We propose a new framework which extracts both global and local information in order to produce proper depth maps. We argue that simple depth completion does not require a deep network. However, we additionally propose a fusion method with RGB guidance from a monocular camera in order to leverage object information and to correct mistakes in the sparse input. This improves the accuracy significantly. Moreover, confidence masks are exploited in order to take into account the uncertainty in the depth predictions from each modality. This fusion method outperforms the state-of-the-art and ranks first on the KITTI depth completion benchmark. Our code with visualizations is available.
Considerable progress has been made in semantic scene understanding of road scenes with monocular cameras. It is, however, mainly related to certain classes such as cars and pedestrians. This work investigates traffic cones, an object class crucial for traffic control in the context of autonomous vehicles. 3D object detection using images from a monocular camera is intrinsically an ill-posed problem. In this work, we leverage the unique structure of traffic cones and propose a pipelined approach to the problem. Specifically, we first detect cones in images by a tailored 2D object detector; then, the spatial arrangement of keypoints on a traffic cone are detected by our deep structural regression network, where the fact that the cross-ratio is projection invariant is leveraged for network regularization; finally, the 3D position of cones is recovered by the classical Perspective n-Point algorithm. Extensive experiments show that our approach can accurately detect traffic cones and estimate their position in the 3D world in real time. The proposed method is also deployed on a real-time, critical system. It runs efficiently on the low-power Jetson TX2, providing accurate 3D position estimates, allowing a race-car to map and drive autonomously on an unseen track indicated by traffic cones. With the help of robust and accurate perception, our race-car won both Formula Student Competitions held in Italy and Germany in 2018, cruising at a top-speed of 54 kmph. Visualization of the complete pipeline, mapping and navigation can be found on our project page.
Lane detection is typically tackled with a two-step pipeline in which a segmentation mask of the lane markings is predicted first, and a lane line model (like a parabola or spline) is fitted to the post-processed mask next. The problem with such a two-step approach is that the parameters of the network are not optimized for the true task of interest (estimating the lane curvature parameters) but for a proxy task (segmenting the lane markings), resulting in sub-optimal performance. In this work, we propose a method to train a lane detector in an end-to-end manner, directly regressing the lane parameters. The architecture consists of two components: a deep network that predicts a segmentation-like weight map for each lane line, and a differentiable least-squares fitting module that returns for each map the parameters of the best-fitting curve in the weighted least-squares sense. These parameters can subsequently be supervised with a loss function of choice. Our method relies on the observation that it is possible to backpropagate through a least-squares fitting procedure. This leads to an end-to-end method where the features are optimized for the true task of interest: the network implicitly learns to generate features that prevent instabilities during the model fitting step, as opposed to two-step pipelines that need to handle outliers with heuristics. Additionally, the system is not just a black box but offers a degree of interpretability because the intermediately generated segmentation-like weight maps can be inspected and visualized. Code and a video is available at github.com/wvangansbeke/LaneDetection_End2End.
This work addresses the problem of semantic segmentation of nighttime images. The main direction of recent progress in semantic segmentation pertains to daytime scenes with favorable illumination conditions. We focus on improving the performance of state-of-the-art methods on the nighttime domain by adapting them to nighttime data without extra annotations, and designing a new evaluation framework to address the uncertainty of semantics in nighttime images. To this end, we make the following contributions: 1) a novel pipeline for dataset-scale guided style transfer to generate synthetic nighttime images from real daytime input; 2) a framework to gradually adapt semantic segmentation models from day to night via stylized and real images of progressively increasing darkness; 3) a novel uncertainty-aware annotation and evaluation framework and metric for semantic segmentation in adverse conditions; 4) the Dark Zurich dataset with 2416 nighttime and 2920 twilight unlabeled images plus 20 nighttime images with pixel-level annotations that conform to our newly-proposed evaluation. Our experiments evidence that both our stylized data per se and our gradual adaptation significantly boost performance at nighttime both for standard evaluation metrics and our metric. Moreover, our new evaluation reveals that state-of-the-art segmentation models output overly confident predictions at indiscernible regions compared to visible ones.
Recently, increasing attention has been drawn to training semantic segmentation models using synthetic data and computer-generated annotation. However, domain gap remains a major barrier and prevents models learned from synthetic data from generalizing well to real-world applications. In this work, we take the advantage of additional geometric information from synthetic data, a powerful yet largely neglected cue, to bridge the domain gap. Such geometric information can be generated easily from synthetic data, and is proven to be closely coupled with semantic information. With the geometric information, we propose a model to reduce domain shift on two levels: on the input level, we augment the traditional image translation network with the additional geometric information to translate synthetic images into realistic styles; on the output level, we build a task network which simultaneously performs depth estimation and semantic segmentation on the synthetic data. Meanwhile, we encourage the network to preserve correlation between depth and semantics by adversarial training on the output space. We then validate our method on two pairs of synthetic to real dataset: Virtual KITTI to KITTI, and SYNTHIA to Cityscapes, where we achieve a significant performance gain compared to the non-adapt baseline and methods using only semantic label. This demonstrates the usefulness of geometric information from synthetic data for cross-domain semantic segmentation.