Antonio M. López

Eugenio Alcalá, Laura Sellart, Vicenç Puig et al.

This paper presents the comparison of two non-linear model-based control strategies for autonomous cars. A control oriented model of vehicle based on a bicycle model is used. The two control strategies use a model reference approach. Using this approach, the error dynamics model is developed. Both controllers receive as input the longitudinal, lateral and orientation errors generating as control outputs the steering angle and the velocity of the vehicle. The first control approach is based on a non-linear control law that is designed by means of the Lyapunov direct approach. The second approach is based on a sliding mode-control that defines a set of sliding surfaces over which the error trajectories will converge. The main advantage of the sliding-control technique is the robustness against non-linearities and parametric uncertainties in the model. However, the main drawback of first order sliding mode is the chattering, so it has been implemented a high order sliding mode control. To test and compare the proposed control strategies, different path following scenarios are used in simulation.

Prassanna Ganesh Ravishankar, Antonio M. Lopez, Gemma M. Sanchez

Monocular vision based road detection methods are mostly based on machine learning methods, relying on classification and feature extraction accuracy, and suffer from appearance, illumination and weather changes. Traditional methods introduce the predictions into conditional random fields or markov random fields models to improve the intermediate predictions based on structure. These methods are optimization based and therefore resource heavy and slow, making it unsuitable for real time applications. We propose a method to detect and segment roads with a random forest classifier of local experts with superpixel based machine-learned features. The random forest takes in machine learnt descriptors from a pre-trained convolutional neural network - VGG-16. The features are also pooled into their respective superpixels, allowing for local structure to be continuous. We compare our algorithm against Nueral Network based methods and Traditional approaches (based on Hand-crafted features), on both Structured Road (CamVid and Kitti) and Unstructured Road Datasets. Finally, we introduce a Road Scene Dataset with 1000 annotated images, and verify that our algorithm works well in non-urban and rural road scenarios.

Akhil Gurram, Antonio M. Lopez

Monocular Depth Estimation (MDE) is performed to produce 3D information that can be used in downstream tasks such as those related to on-board perception for Autonomous Vehicles (AVs) or driver assistance. Therefore, a relevant arising question is whether the standard metrics for MDE assessment are a good indicator of the accuracy of future MDE-based driving-related perception tasks. We address this question in this paper. In particular, we take the task of 3D object detection on point clouds as a proxy of on-board perception. We train and test state-of-the-art 3D object detectors using 3D point clouds coming from MDE models. We confront the ranking of object detection results with the ranking given by the depth estimation metrics of the MDE models. We conclude that, indeed, MDE evaluation metrics give rise to a ranking of methods that reflects relatively well the 3D object detection results we may expect. Among the different metrics, the absolute relative (abs-rel) error seems to be the best for that purpose.

Yi Xiao, Felipe Codevilla, Diego Porres et al.

On end-to-end driving, human driving demonstrations are used to train perception-based driving models by imitation learning. This process is supervised on vehicle signals (e.g., steering angle, acceleration) but does not require extra costly supervision (human labeling of sensor data). As a representative of such vision-based end-to-end driving models, CILRS is commonly used as a baseline to compare with new driving models. So far, some latest models achieve better performance than CILRS by using expensive sensor suites and/or by using large amounts of human-labeled data for training. Given the difference in performance, one may think that it is not worth pursuing vision-based pure end-to-end driving. However, we argue that this approach still has great value and potential considering cost and maintenance. In this paper, we present CIL++, which improves on CILRS by both processing higher-resolution images using a human-inspired HFOV as an inductive bias and incorporating a proper attention mechanism. CIL++ achieves competitive performance compared to models which are more costly to develop. We propose to replace CILRS with CIL++ as a strong vision-based pure end-to-end driving baseline supervised by only vehicle signals and trained by conditional imitation learning.

Maciej Wielgosz, Antonio M. López, Muhammad Naveed Riaz

We present a sample dataset featuring pedestrians generated using the ARCANE framework, a new framework for generating datasets in CARLA (0.9.13). We provide use cases for pedestrian detection, autoencoding, pose estimation, and pose lifting. We also showcase baseline results. For more information, visit https://project-arcane.eu/.

Jose L. Gómez, Gabriel Villalonga, Antonio M. López

Semantic image segmentation is a central and challenging task in autonomous driving, addressed by training deep models. Since this training draws to a curse of human-based image labeling, using synthetic images with automatically generated labels together with unlabeled real-world images is a promising alternative. This implies to address an unsupervised domain adaptation (UDA) problem. In this paper, we propose a new co-training procedure for synth-to-real UDA of semantic segmentation models. It consists of a self-training stage, which provides two domain-adapted models, and a model collaboration loop for the mutual improvement of these two models. These models are then used to provide the final semantic segmentation labels (pseudo-labels) for the real-world images. The overall procedure treats the deep models as black boxes and drives their collaboration at the level of pseudo-labeled target images, i.e., neither modifying loss functions is required, nor explicit feature alignment. We test our proposal on standard synthetic and real-world datasets for on-board semantic segmentation. Our procedure shows improvements ranging from ~13 to ~26 mIoU points over baselines, so establishing new state-of-the-art results.

Yachan Guo, Yi Xiao, Danna Xue et al.

Instance segmentation is crucial for autonomous driving, but is hindered by the lack of annotated real-world data due to expensive labeling costs. Unsupervised Domain Adaptation (UDA) offers a solution by transferring knowledge from labeled synthetic data to unlabeled real-world data. While UDA methods for synthetic to real-world domains (synth-to-real) excel in tasks such as semantic segmentation and object detection, their application to instance segmentation for autonomous driving remains underexplored and often relies on suboptimal baselines. We introduce UDA4Inst, a powerful framework for synth-to-real UDA in instance segmentation. Our framework enhances instance segmentation through Semantic Category Training and Bidirectional Mixing Training. Semantic Category Training groups semantically related classes for separate training, improving pseudo-label quality and segmentation accuracy. Bidirectional Mixing Training combines instance-wise and patch-wise data mixing, creating coherent composites that enhance generalization across domains. Extensive experiments show UDA4Inst sets a new state-of-the-art on the SYNTHIA-> Cityscapes benchmark (mAP 31.3) and introduces results on novel datasets, using UrbanSyn and Synscapes as sources and Cityscapes and KITTI360 as targets. Code and models are available at https://github.com/gyc-code/UDA4Inst.

Jiaolong Xu, Liang Xiao, Antonio M. Lopez

Recent progress of self-supervised visual representation learning has achieved remarkable success on many challenging computer vision benchmarks. However, whether these techniques can be used for domain adaptation has not been explored. In this work, we propose a generic method for self-supervised domain adaptation, using object recognition and semantic segmentation of urban scenes as use cases. Focusing on simple pretext/auxiliary tasks (e.g. image rotation prediction), we assess different learning strategies to improve domain adaptation effectiveness by self-supervision. Additionally, we propose two complementary strategies to further boost the domain adaptation accuracy on semantic segmentation within our method, consisting of prediction layer alignment and batch normalization calibration. The experimental results show adaptation levels comparable to most studied domain adaptation methods, thus, bringing self-supervision as a new alternative for reaching domain adaptation. The code is available at https://github.com/Jiaolong/self-supervised-da.

Felipe Codevilla, Eder Santana, Antonio M. López et al.

Driving requires reacting to a wide variety of complex environment conditions and agent behaviors. Explicitly modeling each possible scenario is unrealistic. In contrast, imitation learning can, in theory, leverage data from large fleets of human-driven cars. Behavior cloning in particular has been successfully used to learn simple visuomotor policies end-to-end, but scaling to the full spectrum of driving behaviors remains an unsolved problem. In this paper, we propose a new benchmark to experimentally investigate the scalability and limitations of behavior cloning. We show that behavior cloning leads to state-of-the-art results, including in unseen environments, executing complex lateral and longitudinal maneuvers without these reactions being explicitly programmed. However, we confirm well-known limitations (due to dataset bias and overfitting), new generalization issues (due to dynamic objects and the lack of a causal model), and training instability requiring further research before behavior cloning can graduate to real-world driving. The code of the studied behavior cloning approaches can be found at https://github.com/felipecode/coiltraine .

Alexey Dosovitskiy, German Ros, Felipe Codevilla et al.

We introduce CARLA, an open-source simulator for autonomous driving research. CARLA has been developed from the ground up to support development, training, and validation of autonomous urban driving systems. In addition to open-source code and protocols, CARLA provides open digital assets (urban layouts, buildings, vehicles) that were created for this purpose and can be used freely. The simulation platform supports flexible specification of sensor suites and environmental conditions. We use CARLA to study the performance of three approaches to autonomous driving: a classic modular pipeline, an end-to-end model trained via imitation learning, and an end-to-end model trained via reinforcement learning. The approaches are evaluated in controlled scenarios of increasing difficulty, and their performance is examined via metrics provided by CARLA, illustrating the platform's utility for autonomous driving research. The supplementary video can be viewed at https://youtu.be/Hp8Dz-Zek2E

Jose L. Gómez, Manuel Silva, Antonio Seoane et al.

We introduce UrbanSyn, a photorealistic dataset acquired through semi-procedurally generated synthetic urban driving scenarios. Developed using high-quality geometry and materials, UrbanSyn provides pixel-level ground truth, including depth, semantic segmentation, and instance segmentation with object bounding boxes and occlusion degree. It complements GTAV and Synscapes datasets to form what we coin as the 'Three Musketeers'. We demonstrate the value of the Three Musketeers in unsupervised domain adaptation for image semantic segmentation. Results on real-world datasets, Cityscapes, Mapillary Vistas, and BDD100K, establish new benchmarks, largely attributed to UrbanSyn. We make UrbanSyn openly and freely accessible (www.urbansyn.org).

Muhammad Naveed Riaz, Maciej Wielgosz, Abel Garcia Romera et al.

Pedestrian intention prediction is crucial for autonomous driving. In particular, knowing if pedestrians are going to cross in front of the ego-vehicle is core to performing safe and comfortable maneuvers. Creating accurate and fast models that predict such intentions from sequential images is challenging. A factor contributing to this is the lack of datasets with diverse crossing and non-crossing (C/NC) scenarios. We address this scarceness by introducing a framework, named ARCANE, which allows programmatically generating synthetic datasets consisting of C/NC video clip samples. As an example, we use ARCANE to generate a large and diverse dataset named PedSynth. We will show how PedSynth complements widely used real-world datasets such as JAAD and PIE, so enabling more accurate models for C/NC prediction. Considering the onboard deployment of C/NC prediction models, we also propose a deep model named PedGNN, which is fast and has a very low memory footprint. PedGNN is based on a GNN-GRU architecture that takes a sequence of pedestrian skeletons as input to predict crossing intentions.

Antonio Lopez, David Fridovich-Keil

Recent methods using Reinforcement Learning (RL) have proven to be successful for training intelligent agents in unknown environments. However, RL has not been applied widely in real-world robotics scenarios. This is because current state-of-the-art RL methods require large amounts of data to learn a specific task, leading to unreasonable costs when deploying the agent to collect data in real-world applications. In this paper, we build from existing work that reshapes the reward function in RL by introducing a Control Lyapunov Function (CLF), which is demonstrated to reduce the sample complexity. Still, this formulation requires knowing a CLF of the system, but due to the lack of a general method, it is often a challenge to identify a suitable CLF. Existing work can compute low-dimensional CLFs via a Hamilton-Jacobi reachability procedure. However, this class of methods becomes intractable on high-dimensional systems, a problem that we address by using a system decomposition technique to compute what we call Decomposed Control Lyapunov Functions (DCLFs). We use the computed DCLF for reward shaping, which we show improves RL performance. Through multiple examples, we demonstrate the effectiveness of this approach, where our method finds a policy to successfully land a quadcopter in less than half the amount of real-world data required by the state-of-the-art Soft-Actor Critic algorithm.

Diego Porres, Yi Xiao, Gabriel Villalonga et al.

Vision-based end-to-end driving models trained by imitation learning can lead to affordable solutions for autonomous driving. However, training these well-performing models usually requires a huge amount of data, while still lacking explicit and intuitive activation maps to reveal the inner workings of these models while driving. In this paper, we study how to guide the attention of these models to improve their driving quality and obtain more intuitive activation maps by adding a loss term during training using salient semantic maps. In contrast to previous work, our method does not require these salient semantic maps to be available during testing time, as well as removing the need to modify the model's architecture to which it is applied. We perform tests using perfect and noisy salient semantic maps with encouraging results in both, the latter of which is inspired by possible errors encountered with real data. Using CIL++ as a representative state-of-the-art model and the CARLA simulator with its standard benchmarks, we conduct experiments that show the effectiveness of our method in training better autonomous driving models, especially when data and computational resources are scarce.

Tao Lian, Jose L. Gómez, Antonio M. López

The last mile of unsupervised domain adaptation (UDA) for semantic segmentation is the challenge of solving the syn-to-real domain gap. Recent UDA methods have progressed significantly, yet they often rely on strategies customized for synthetic single-source datasets (e.g., GTA5), which limits their generalisation to multi-source datasets. Conversely, synthetic multi-source datasets hold promise for advancing the last mile of UDA but remain underutilized in current research. Thus, we propose DEC, a flexible UDA framework for multi-source datasets. Following a divide-and-conquer strategy, DEC simplifies the task by categorizing semantic classes, training models for each category, and fusing their outputs by an ensemble model trained exclusively on synthetic datasets to obtain the final segmentation mask. DEC can integrate with existing UDA methods, achieving state-of-the-art performance on Cityscapes, BDD100K, and Mapillary Vistas, significantly narrowing the syn-to-real domain gap.

Daniel Coelho, Miguel Oliveira, Vitor Santos et al.

The development of Autonomous Driving (AD) systems in simulated environments like CARLA is crucial for advancing real-world automotive technologies. To drive innovation, CARLA introduced Leaderboard 2.0, significantly more challenging than its predecessor. However, current AD methods have struggled to achieve satisfactory outcomes due to a lack of sufficient ground truth data. Human driving logs provided by CARLA are insufficient, and previously successful expert agents like Autopilot and Roach, used for collecting datasets, have seen reduced effectiveness under these more demanding conditions. To overcome these data limitations, we introduce PRIBOOT, an expert agent that leverages limited human logs with privileged information. We have developed a novel BEV representation specifically tailored to meet the demands of this new benchmark and processed it as an RGB image to facilitate the application of transfer learning techniques, instead of using a set of masks. Additionally, we propose the Infraction Rate Score (IRS), a new evaluation metric designed to provide a more balanced assessment of driving performance over extended routes. PRIBOOT is the first model to achieve a Route Completion (RC) of 75% in Leaderboard 2.0, along with a Driving Score (DS) and IRS of 20% and 45%, respectively. With PRIBOOT, researchers can now generate extensive datasets, potentially solving the data availability issues that have hindered progress in this benchmark.

Yufan Zhu, Chongzhi Ran, Mingtao Feng et al.

Accurate monocular depth estimation is a fundamental component of vision-based perception systems in intelligent transportation applications. Despite recent progress, unsupervised monocular approaches still suffer from significant performance degradation in real-world traffic scenes due to synthetic-to-real domain gaps and the presence of dynamic, non-rigid objects such as vehicles and pedestrians. In this paper, we propose Back2Color, a robust unsupervised monocular depth estimation framework that addresses these challenges through domain adaptation and uncertainty-aware fusion. Specifically, Back2Color proposes a bidirectional depth-to-color transformation strategy that learns appearance mappings from real-world driving data and applies them to synthetic depth maps, thereby constructing training samples with realistic color appearance and paired synthetic depth. In this way, the proposed approach effectively reduces the domain gap between simulated and real traffic scenes, enabling the depth prediction network to learn more stable and generalizable priors. To further improve robustness under dynamic environments, we propose an auto-learning uncertainty temporal-spatial fusion (Auto-UTSF) module, which adaptively fuses complementary temporal and spatial cues by estimating pixel-wise uncertainty, enabling reliable depth prediction in the presence of moving objects and occlusions. Extensive experiments on challenging urban driving benchmarks, including KITTI and Cityscapes, demonstrate that the proposed method consistently outperforms existing unsupervised monocular depth estimation approaches, particularly in dynamic traffic scenarios, while maintaining high computational efficiency.

Akhil Gurram, Ahmet Faruk Tuna, Fengyi Shen et al.

Depth information is essential for on-board perception in autonomous driving and driver assistance. Monocular depth estimation (MDE) is very appealing since it allows for appearance and depth being on direct pixelwise correspondence without further calibration. Best MDE models are based on Convolutional Neural Networks (CNNs) trained in a supervised manner, i.e., assuming pixelwise ground truth (GT). Usually, this GT is acquired at training time through a calibrated multi-modal suite of sensors. However, also using only a monocular system at training time is cheaper and more scalable. This is possible by relying on structure-from-motion (SfM) principles to generate self-supervision. Nevertheless, problems of camouflaged objects, visibility changes, static-camera intervals, textureless areas, and scale ambiguity, diminish the usefulness of such self-supervision. In this paper, we perform monocular depth estimation by virtual-world supervision (MonoDEVS) and real-world SfM self-supervision. We compensate the SfM self-supervision limitations by leveraging virtual-world images with accurate semantic and depth supervision and addressing the virtual-to-real domain gap. Our MonoDEVSNet outperforms previous MDE CNNs trained on monocular and even stereo sequences.

Yi Xiao, Felipe Codevilla, Christopher Pal et al.

Human drivers produce a vast amount of data which could, in principle, be used to improve autonomous driving systems. Unfortunately, seemingly straightforward approaches for creating end-to-end driving models that map sensor data directly into driving actions are problematic in terms of interpretability, and typically have significant difficulty dealing with spurious correlations. Alternatively, we propose to use this kind of action-based driving data for learning representations. Our experiments show that an affordance-based driving model pre-trained with this approach can leverage a relatively small amount of weakly annotated imagery and outperform pure end-to-end driving models, while being more interpretable. Further, we demonstrate how this strategy outperforms previous methods based on learning inverse dynamics models as well as other methods based on heavy human supervision (ImageNet).

Gabriel Villalonga, Antonio M. Lopez

Providing ground truth supervision to train visual models has been a bottleneck over the years, exacerbated by domain shifts which degenerate the performance of such models. This was the case when visual tasks relied on handcrafted features and shallow machine learning and, despite its unprecedented performance gains, the problem remains open within the deep learning paradigm due to its data-hungry nature. Best performing deep vision-based object detectors are trained in a supervised manner by relying on human-labeled bounding boxes which localize class instances (i.e.objects) within the training images.Thus, object detection is one of such tasks for which human labeling is a major bottleneck. In this paper, we assess co-training as a semi-supervised learning method for self-labeling objects in unlabeled images, so reducing the human-labeling effort for developing deep object detectors. Our study pays special attention to a scenario involving domain shift; in particular, when we have automatically generated virtual-world images with object bounding boxes and we have real-world images which are unlabeled. Moreover, we are particularly interested in using co-training for deep object detection in the context of driver assistance systems and/or self-driving vehicles. Thus, using well-established datasets and protocols for object detection in these application contexts, we will show how co-training is a paradigm worth to pursue for alleviating object labeling, working both alone and together with task-agnostic domain adaptation.

Sudeep Katakol, Basem Elbarashy, Luis Herranz et al.

Modern computer vision requires processing large amounts of data, both while training the model and/or during inference, once the model is deployed. Scenarios where images are captured and processed in physically separated locations are increasingly common (e.g. autonomous vehicles, cloud computing). In addition, many devices suffer from limited resources to store or transmit data (e.g. storage space, channel capacity). In these scenarios, lossy image compression plays a crucial role to effectively increase the number of images collected under such constraints. However, lossy compression entails some undesired degradation of the data that may harm the performance of the downstream analysis task at hand, since important semantic information may be lost in the process. Moreover, we may only have compressed images at training time but are able to use original images at inference time, or vice versa, and in such a case, the downstream model suffers from covariate shift. In this paper, we analyze this phenomenon, with a special focus on vision-based perception for autonomous driving as a paradigmatic scenario. We see that loss of semantic information and covariate shift do indeed exist, resulting in a drop in performance that depends on the compression rate. In order to address the problem, we propose dataset restoration, based on image restoration with generative adversarial networks (GANs). Our method is agnostic to both the particular image compression method and the downstream task; and has the advantage of not adding additional cost to the deployed models, which is particularly important in resource-limited devices. The presented experiments focus on semantic segmentation as a challenging use case, cover a broad range of compression rates and diverse datasets, and show how our method is able to significantly alleviate the negative effects of compression on the downstream visual task.

Hamed H. Aghdam, Abel Gonzalez-Garcia, Joost van de Weijer et al.

The cost of drawing object bounding boxes (i.e. labeling) for millions of images is prohibitively high. For instance, labeling pedestrians in a regular urban image could take 35 seconds on average. Active learning aims to reduce the cost of labeling by selecting only those images that are informative to improve the detection network accuracy. In this paper, we propose a method to perform active learning of object detectors based on convolutional neural networks. We propose a new image-level scoring process to rank unlabeled images for their automatic selection, which clearly outperforms classical scores. The proposed method can be applied to videos and sets of still images. In the former case, temporal selection rules can complement our scoring process. As a relevant use case, we extensively study the performance of our method on the task of pedestrian detection. Overall, the experiments show that the proposed method performs better than random selection. Our codes are publicly available at www.gitlab.com/haghdam/deep_active_learning.

César Roberto de Souza, Adrien Gaidon, Yohann Cabon et al.

Deep video action recognition models have been highly successful in recent years but require large quantities of manually annotated data, which are expensive and laborious to obtain. In this work, we investigate the generation of synthetic training data for video action recognition, as synthetic data have been successfully used to supervise models for a variety of other computer vision tasks. We propose an interpretable parametric generative model of human action videos that relies on procedural generation, physics models and other components of modern game engines. With this model we generate a diverse, realistic, and physically plausible dataset of human action videos, called PHAV for "Procedural Human Action Videos". PHAV contains a total of 39,982 videos, with more than 1,000 examples for each of 35 action categories. Our video generation approach is not limited to existing motion capture sequences: 14 of these 35 categories are procedurally defined synthetic actions. In addition, each video is represented with 6 different data modalities, including RGB, optical flow and pixel-level semantic labels. These modalities are generated almost simultaneously using the Multiple Render Targets feature of modern GPUs. In order to leverage PHAV, we introduce a deep multi-task (i.e. that considers action classes from multiple datasets) representation learning architecture that is able to simultaneously learn from synthetic and real video datasets, even when their action categories differ. Our experiments on the UCF-101 and HMDB-51 benchmarks suggest that combining our large set of synthetic videos with small real-world datasets can boost recognition performance. Our approach also significantly outperforms video representations produced by fine-tuning state-of-the-art unsupervised generative models of videos.

Zhijie Fang, Antonio M. López

Anticipating the intentions of vulnerable road users (VRUs) such as pedestrians and cyclists is critical for performing safe and comfortable driving maneuvers. This is the case for human driving and, thus, should be taken into account by systems providing any level of driving assistance, from advanced driver assistant systems (ADAS) to fully autonomous vehicles (AVs). In this paper, we show how the latest advances on monocular vision-based human pose estimation, i.e. those relying on deep Convolutional Neural Networks (CNNs), enable to recognize the intentions of such VRUs. In the case of cyclists, we assume that they follow traffic rules to indicate future maneuvers with arm signals. In the case of pedestrians, no indications can be assumed. Instead, we hypothesize that the walking pattern of a pedestrian allows to determine if he/she has the intention of crossing the road in the path of the ego-vehicle, so that the ego-vehicle must maneuver accordingly (e.g. slowing down or stopping). In this paper, we show how the same methodology can be used for recognizing pedestrians and cyclists' intentions. For pedestrians, we perform experiments on the JAAD dataset. For cyclists, we did not found an analogous dataset, thus, we created our own one by acquiring and annotating videos which we share with the research community. Overall, the proposed pipeline provides new state-of-the-art results on the intention recognition of VRUs.

Daniel Hernandez-Juarez, Lukas Schneider, Pau Cebrian et al.

This work presents and evaluates a novel compact scene representation based on Stixels that infers geometric and semantic information. Our approach overcomes the previous rather restrictive geometric assumptions for Stixels by introducing a novel depth model to account for non-flat roads and slanted objects. Both semantic and depth cues are used jointly to infer the scene representation in a sound global energy minimization formulation. Furthermore, a novel approximation scheme is introduced in order to significantly reduce the computational complexity of the Stixel algorithm, and then achieve real-time computation capabilities. The idea is to first perform an over-segmentation of the image, discarding the unlikely Stixel cuts, and apply the algorithm only on the remaining Stixel cuts. This work presents a novel over-segmentation strategy based on a Fully Convolutional Network (FCN), which outperforms an approach based on using local extrema of the disparity map. We evaluate the proposed methods in terms of semantic and geometric accuracy as well as run-time on four publicly available benchmark datasets. Our approach maintains accuracy on flat road scene datasets while improving substantially on a novel non-flat road dataset.

Javad Zolfaghari Bengar, Abel Gonzalez-Garcia, Gabriel Villalonga et al.

Autonomous driving systems require huge amounts of data to train. Manual annotation of this data is time-consuming and prohibitively expensive since it involves human resources. Therefore, active learning emerged as an alternative to ease this effort and to make data annotation more manageable. In this paper, we introduce a novel active learning approach for object detection in videos by exploiting temporal coherence. Our active learning criterion is based on the estimated number of errors in terms of false positives and false negatives. The detections obtained by the object detector are used to define the nodes of a graph and tracked forward and backward to temporally link the nodes. Minimizing an energy function defined on this graphical model provides estimates of both false positives and false negatives. Additionally, we introduce a synthetic video dataset, called SYNTHIA-AL, specially designed to evaluate active learning for video object detection in road scenes. Finally, we show that our approach outperforms active learning baselines tested on two datasets.

Yi Xiao, Felipe Codevilla, Akhil Gurram et al.

A crucial component of an autonomous vehicle (AV) is the artificial intelligence (AI) is able to drive towards a desired destination. Today, there are different paradigms addressing the development of AI drivers. On the one hand, we find modular pipelines, which divide the driving task into sub-tasks such as perception and maneuver planning and control. On the other hand, we find end-to-end driving approaches that try to learn a direct mapping from input raw sensor data to vehicle control signals. The later are relatively less studied, but are gaining popularity since they are less demanding in terms of sensor data annotation. This paper focuses on end-to-end autonomous driving. So far, most proposals relying on this paradigm assume RGB images as input sensor data. However, AVs will not be equipped only with cameras, but also with active sensors providing accurate depth information (e.g., LiDARs). Accordingly, this paper analyses whether combining RGB and depth modalities, i.e. using RGBD data, produces better end-to-end AI drivers than relying on a single modality. We consider multimodality based on early, mid and late fusion schemes, both in multisensory and single-sensor (monocular depth estimation) settings. Using the CARLA simulator and conditional imitation learning (CIL), we show how, indeed, early fusion multimodality outperforms single-modality.

Felipe Codevilla, Antonio M. López, Vladlen Koltun et al.

Autonomous driving models should ideally be evaluated by deploying them on a fleet of physical vehicles in the real world. Unfortunately, this approach is not practical for the vast majority of researchers. An attractive alternative is to evaluate models offline, on a pre-collected validation dataset with ground truth annotation. In this paper, we investigate the relation between various online and offline metrics for evaluation of autonomous driving models. We find that offline prediction error is not necessarily correlated with driving quality, and two models with identical prediction error can differ dramatically in their driving performance. We show that the correlation of offline evaluation with driving quality can be significantly improved by selecting an appropriate validation dataset and suitable offline metrics. The supplementary video can be viewed at https://www.youtube.com/watch?v=P8K8Z-iF0cY

Marc Masana, Idoia Ruiz, Joan Serrat et al.

When neural networks process images which do not resemble the distribution seen during training, so called out-of-distribution images, they often make wrong predictions, and do so too confidently. The capability to detect out-of-distribution images is therefore crucial for many real-world applications. We divide out-of-distribution detection between novelty detection ---images of classes which are not in the training set but are related to those---, and anomaly detection ---images with classes which are unrelated to the training set. By related we mean they contain the same type of objects, like digits in MNIST and SVHN. Most existing work has focused on anomaly detection, and has addressed this problem considering networks trained with the cross-entropy loss. Differently from them, we propose to use metric learning which does not have the drawback of the softmax layer (inherent to cross-entropy methods), which forces the network to divide its prediction power over the learned classes. We perform extensive experiments and evaluate both novelty and anomaly detection, even in a relevant application such as traffic sign recognition, obtaining comparable or better results than previous works.

Zhijie Fang, Antonio M. López

Our recent work suggests that, thanks to nowadays powerful CNNs, image-based 2D pose estimation is a promising cue for determining pedestrian intentions such as crossing the road in the path of the ego-vehicle, stopping before entering the road, and starting to walk or bending towards the road. This statement is based on the results obtained on non-naturalistic sequences (Daimler dataset), i.e. in sequences choreographed specifically for performing the study. Fortunately, a new publicly available dataset (JAAD) has appeared recently to allow developing methods for detecting pedestrian intentions in naturalistic driving conditions; more specifically, for addressing the relevant question is the pedestrian going to cross? Accordingly, in this paper we use JAAD to assess the usefulness of 2D pose estimation for answering such a question. We combine CNN-based pedestrian detection, tracking and pose estimation to predict the crossing action from monocular images. Overall, the proposed pipeline provides new state-of-the-art results.

Jiaolong Xu, Peng Wang, Heng Yang et al.

Autonomous driving has harsh requirements of small model size and energy efficiency, in order to enable the embedded system to achieve real-time on-board object detection. Recent deep convolutional neural network based object detectors have achieved state-of-the-art accuracy. However, such models are trained with numerous parameters and their high computational costs and large storage prohibit the deployment to memory and computation resource limited systems. Low-precision neural networks are popular techniques for reducing the computation requirements and memory footprint. Among them, binary weight neural network (BWN) is the extreme case which quantizes the float-point into just $1$ bit. BWNs are difficult to train and suffer from accuracy deprecation due to the extreme low-bit representation. To address this problem, we propose a knowledge transfer (KT) method to aid the training of BWN using a full-precision teacher network. We built DarkNet- and MobileNet-based binary weight YOLO-v2 detectors and conduct experiments on KITTI benchmark for car, pedestrian and cyclist detection. The experimental results show that the proposed method maintains high detection accuracy while reducing the model size of DarkNet-YOLO from 257 MB to 8.8 MB and MobileNet-YOLO from 193 MB to 7.9 MB.

Akhil Gurram, Onay Urfalioglu, Ibrahim Halfaoui et al.

Depth estimation provides essential information to perform autonomous driving and driver assistance. Especially, Monocular Depth Estimation is interesting from a practical point of view, since using a single camera is cheaper than many other options and avoids the need for continuous calibration strategies as required by stereo-vision approaches. State-of-the-art methods for Monocular Depth Estimation are based on Convolutional Neural Networks (CNNs). A promising line of work consists of introducing additional semantic information about the traffic scene when training CNNs for depth estimation. In practice, this means that the depth data used for CNN training is complemented with images having pixel-wise semantic labels, which usually are difficult to annotate (e.g. crowded urban images). Moreover, so far it is common practice to assume that the same raw training data is associated with both types of ground truth, i.e., depth and semantic labels. The main contribution of this paper is to show that this hard constraint can be circumvented, i.e., that we can train CNNs for depth estimation by leveraging the depth and semantic information coming from heterogeneous datasets. In order to illustrate the benefits of our approach, we combine KITTI depth and Cityscapes semantic segmentation datasets, outperforming state-of-the-art results on Monocular Depth Estimation.

Xialei Liu, Marc Masana, Luis Herranz et al.

In this paper we propose an approach to avoiding catastrophic forgetting in sequential task learning scenarios. Our technique is based on a network reparameterization that approximately diagonalizes the Fisher Information Matrix of the network parameters. This reparameterization takes the form of a factorized rotation of parameter space which, when used in conjunction with Elastic Weight Consolidation (which assumes a diagonal Fisher Information Matrix), leads to significantly better performance on lifelong learning of sequential tasks. Experimental results on the MNIST, CIFAR-100, CUB-200 and Stanford-40 datasets demonstrate that we significantly improve the results of standard elastic weight consolidation, and that we obtain competitive results when compared to other state-of-the-art in lifelong learning without forgetting.

Felipe Codevilla, Matthias Müller, Antonio López et al.

Deep networks trained on demonstrations of human driving have learned to follow roads and avoid obstacles. However, driving policies trained via imitation learning cannot be controlled at test time. A vehicle trained end-to-end to imitate an expert cannot be guided to take a specific turn at an upcoming intersection. This limits the utility of such systems. We propose to condition imitation learning on high-level command input. At test time, the learned driving policy functions as a chauffeur that handles sensorimotor coordination but continues to respond to navigational commands. We evaluate different architectures for conditional imitation learning in vision-based driving. We conduct experiments in realistic three-dimensional simulations of urban driving and on a 1/5 scale robotic truck that is trained to drive in a residential area. Both systems drive based on visual input yet remain responsive to high-level navigational commands. The supplementary video can be viewed at https://youtu.be/cFtnflNe5fM

Daniel Hernandez-Juarez, Lukas Schneider, Antonio Espinosa et al.

In this work we present a novel compact scene representation based on Stixels that infers geometric and semantic information. Our approach overcomes the previous rather restrictive geometric assumptions for Stixels by introducing a novel depth model to account for non-flat roads and slanted objects. Both semantic and depth cues are used jointly to infer the scene representation in a sound global energy minimization formulation. Furthermore, a novel approximation scheme is introduced that uses an extremely efficient over-segmentation. In doing so, the computational complexity of the Stixel inference algorithm is reduced significantly, achieving real-time computation capabilities with only a slight drop in accuracy. We evaluate the proposed approach in terms of semantic and geometric accuracy as well as run-time on four publicly available benchmark datasets. Our approach maintains accuracy on flat road scene datasets while improving substantially on a novel non-flat road dataset.

Antonio M. Lopez, Jiaolong Xu, Jose L. Gomez et al.

Supervised learning tends to produce more accurate classifiers than unsupervised learning in general. This implies that training data is preferred with annotations. When addressing visual perception challenges, such as localizing certain object classes within an image, the learning of the involved classifiers turns out to be a practical bottleneck. The reason is that, at least, we have to frame object examples with bounding boxes in thousands of images. A priori, the more complex the model is regarding its number of parameters, the more annotated examples are required. This annotation task is performed by human oracles, which ends up in inaccuracies and errors in the annotations (aka ground truth) since the task is inherently very cumbersome and sometimes ambiguous. As an alternative we have pioneered the use of virtual worlds for collecting such annotations automatically and with high precision. However, since the models learned with virtual data must operate in the real world, we still need to perform domain adaptation (DA). In this chapter we revisit the DA of a deformable part-based model (DPM) as an exemplifying case of virtual- to-real-world DA. As a use case, we address the challenge of vehicle detection for driver assistance, using different publicly available virtual-world data. While doing so, we investigate questions such as: how does the domain gap behave due to virtual-vs-real data with respect to dominant object appearance per domain, as well as the role of photo-realism in the virtual world.

César Roberto de Souza, Adrien Gaidon, Yohann Cabon et al.

Deep learning for human action recognition in videos is making significant progress, but is slowed down by its dependency on expensive manual labeling of large video collections. In this work, we investigate the generation of synthetic training data for action recognition, as it has recently shown promising results for a variety of other computer vision tasks. We propose an interpretable parametric generative model of human action videos that relies on procedural generation and other computer graphics techniques of modern game engines. We generate a diverse, realistic, and physically plausible dataset of human action videos, called PHAV for "Procedural Human Action Videos". It contains a total of 39,982 videos, with more than 1,000 examples for each action of 35 categories. Our approach is not limited to existing motion capture sequences, and we procedurally define 14 synthetic actions. We introduce a deep multi-task representation learning architecture to mix synthetic and real videos, even if the action categories differ. Our experiments on the UCF101 and HMDB51 benchmarks suggest that combining our large set of synthetic videos with small real-world datasets can boost recognition performance, significantly outperforming fine-tuning state-of-the-art unsupervised generative models of videos.

David Vázquez, Jorge Bernal, F. Javier Sánchez et al.

Colorectal cancer (CRC) is the third cause of cancer death worldwide. Currently, the standard approach to reduce CRC-related mortality is to perform regular screening in search for polyps and colonoscopy is the screening tool of choice. The main limitations of this screening procedure are polyp miss-rate and inability to perform visual assessment of polyp malignancy. These drawbacks can be reduced by designing Decision Support Systems (DSS) aiming to help clinicians in the different stages of the procedure by providing endoluminal scene segmentation. Thus, in this paper, we introduce an extended benchmark of colonoscopy image, with the hope of establishing a new strong benchmark for colonoscopy image analysis research. We provide new baselines on this dataset by training standard fully convolutional networks (FCN) for semantic segmentation and significantly outperforming, without any further post-processing, prior results in endoluminal scene segmentation.

Azadeh S. Mozafari, David Vazquez, Mansour Jamzad et al.

Random Forest (RF) is a successful paradigm for learning classifiers due to its ability to learn from large feature spaces and seamlessly integrate multi-class classification, as well as the achieved accuracy and processing efficiency. However, as many other classifiers, RF requires domain adaptation (DA) provided that there is a mismatch between the training (source) and testing (target) domains which provokes classification degradation. Consequently, different RF-DA methods have been proposed, which not only require target-domain samples but revisiting the source-domain ones, too. As novelty, we propose three inherently different methods (Node-Adapt, Path-Adapt and Tree-Adapt) that only require the learned source-domain RF and a relatively few target-domain samples for DA, i.e. source-domain samples do not need to be available. To assess the performance of our proposals we focus on image-based object detection, using the pedestrian detection problem as challenging proof-of-concept. Moreover, we use the RF with expert nodes because it is a competitive patch-based pedestrian model. We test our Node-, Path- and Tree-Adapt methods in standard benchmarks, showing that DA is largely achieved.

Victor Campmany, Sergio Silva, Antonio Espinosa et al.

We propose a real-time pedestrian detection system for the embedded Nvidia Tegra X1 GPU-CPU hybrid platform. The pipeline is composed by the following state-of-the-art algorithms: Histogram of Local Binary Patterns (LBP) and Histograms of Oriented Gradients (HOG) features extracted from the input image; Pyramidal Sliding Window technique for candidate generation; and Support Vector Machine (SVM) for classification. Results show a 8x speedup in the target Tegra X1 platform and a better performance/watt ratio than desktop CUDA platforms in study.

Daniel Hernandez-Juarez, Antonio Espinosa, David Vázquez et al.

The Stixel World is a medium-level, compact representation of road scenes that abstracts millions of disparity pixels into hundreds or thousands of stixels. The goal of this work is to implement and evaluate a complete multi-stixel estimation pipeline on an embedded, energy-efficient, GPU-accelerated device. This work presents a full GPU-accelerated implementation of stixel estimation that produces reliable results at 26 frames per second (real-time) on the Tegra X1 for disparity images of 1024x440 pixels and stixel widths of 5 pixels, and achieves more than 400 frames per second on a high-end Titan X GPU card.

Daniel Hernandez-Juarez, Alejandro Chacón, Antonio Espinosa et al.

Dense, robust and real-time computation of depth information from stereo-camera systems is a computationally demanding requirement for robotics, advanced driver assistance systems (ADAS) and autonomous vehicles. Semi-Global Matching (SGM) is a widely used algorithm that propagates consistency constraints along several paths across the image. This work presents a real-time system producing reliable disparity estimation results on the new embedded energy-efficient GPU devices. Our design runs on a Tegra X1 at 42 frames per second (fps) for an image size of 640x480, 128 disparity levels, and using 4 path directions for the SGM method.

César Roberto de Souza, Adrien Gaidon, Eleonora Vig et al.

Action recognition in videos is a challenging task due to the complexity of the spatio-temporal patterns to model and the difficulty to acquire and learn on large quantities of video data. Deep learning, although a breakthrough for image classification and showing promise for videos, has still not clearly superseded action recognition methods using hand-crafted features, even when training on massive datasets. In this paper, we introduce hybrid video classification architectures based on carefully designed unsupervised representations of hand-crafted spatio-temporal features classified by supervised deep networks. As we show in our experiments on five popular benchmarks for action recognition, our hybrid model combines the best of both worlds: it is data efficient (trained on 150 to 10000 short clips) and yet improves significantly on the state of the art, including recent deep models trained on millions of manually labelled images and videos.

Jose M. Alvarez, Theo Gevers, Antonio M. Lopez

Detecting traversable road areas ahead a moving vehicle is a key process for modern autonomous driving systems. A common approach to road detection consists of exploiting color features to classify pixels as road or background. These algorithms reduce the effect of lighting variations and weather conditions by exploiting the discriminant/invariant properties of different color representations. Furthermore, the lack of labeled datasets has motivated the development of algorithms performing on single images based on the assumption that the bottom part of the image belongs to the road surface. In this paper, we first introduce a dataset of road images taken at different times and in different scenarios using an onboard camera. Then, we devise a simple online algorithm and conduct an exhaustive evaluation of different classifiers and the effect of using different color representation to characterize pixels.

José M. Álvarez, Ferran Diego, Joan Serrat et al.

Vision-based road detection is an essential functionality for supporting advanced driver assistance systems (ADAS) such as road following and vehicle and pedestrian detection. The major challenges of road detection are dealing with shadows and lighting variations and the presence of other objects in the scene. Current road detection algorithms characterize road areas at pixel level and group pixels accordingly. However, these algorithms fail in presence of strong shadows and lighting variations. Therefore, we propose a road detection algorithm based on video alignment. The key idea of the algorithm is to exploit the similarities occurred when a vehicle follows the same trajectory more than once. In this way, road areas are learned in a first ride and then, this road knowledge is used to infer areas depicting drivable road surfaces in subsequent rides. Two different experiments are conducted to validate the proposal on different video sequences taken at different scenarios and different daytime. The former aims to perform on-line road detection. The latter aims to perform off-line road detection and is applied to automatically generate the ground-truth necessary to validate road detection algorithms. Qualitative and quantitative evaluations prove that the proposed algorithm is a valid road detection approach.

Jiaolong Xu, Sebastian Ramos, David Vazquez et al.

A key topic in classification is the accuracy loss produced when the data distribution in the training (source) domain differs from that in the testing (target) domain. This is being recognized as a very relevant problem for many computer vision tasks such as image classification, object detection, and object category recognition. In this paper, we present a novel domain adaptation method that leverages multiple target domains (or sub-domains) in a hierarchical adaptation tree. The core idea is to exploit the commonalities and differences of the jointly considered target domains. Given the relevance of structural SVM (SSVM) classifiers, we apply our idea to the adaptive SSVM (A-SSVM), which only requires the target domain samples together with the existing source-domain classifier for performing the desired adaptation. Altogether, we term our proposal as hierarchical A-SSVM (HA-SSVM). As proof of concept we use HA-SSVM for pedestrian detection and object category recognition. In the former we apply HA-SSVM to the deformable part-based model (DPM) while in the latter HA-SSVM is applied to multi-category classifiers. In both cases, we show how HA-SSVM is effective in increasing the detection/recognition accuracy with respect to adaptation strategies that ignore the structure of the target data. Since, the sub-domains of the target data are not always known a priori, we shown how HA-SSVM can incorporate sub-domain structure discovery for object category recognition.

Alejandro González, Sebastian Ramos, David Vázquez et al.

Pedestrian classifiers decide which image windows contain a pedestrian. In practice, such classifiers provide a relatively high response at neighbor windows overlapping a pedestrian, while the responses around potential false positives are expected to be lower. An analogous reasoning applies for image sequences. If there is a pedestrian located within a frame, the same pedestrian is expected to appear close to the same location in neighbor frames. Therefore, such a location has chances of receiving high classification scores during several frames, while false positives are expected to be more spurious. In this paper we propose to exploit such correlations for improving the accuracy of base pedestrian classifiers. In particular, we propose to use two-stage classifiers which not only rely on the image descriptors required by the base classifiers but also on the response of such base classifiers in a given spatiotemporal neighborhood. More specifically, we train pedestrian classifiers using a stacked sequential learning (SSL) paradigm. We use a new pedestrian dataset we have acquired from a car to evaluate our proposal at different frame rates. We also test on a well known dataset: Caltech. The obtained results show that our SSL proposal boosts detection accuracy significantly with a minimal impact on the computational cost. Interestingly, SSL improves more the accuracy at the most dangerous situations, i.e. when a pedestrian is close to the camera.