Self-driving cars must detect other vehicles and pedestrians in 3D to plan safe routes and avoid collisions. State-of-the-art 3D object detectors, based on deep learning, have shown promising accuracy but are prone to over-fit to domain idiosyncrasies, making them fail in new environments -- a serious problem if autonomous vehicles are meant to operate freely. In this paper, we propose a novel learning approach that drastically reduces this gap by fine-tuning the detector on pseudo-labels in the target domain, which our method generates while the vehicle is parked, based on replays of previously recorded driving sequences. In these replays, objects are tracked over time, and detections are interpolated and extrapolated -- crucially, leveraging future information to catch hard cases. We show, on five autonomous driving datasets, that fine-tuning the object detector on these pseudo-labels substantially reduces the domain gap to new driving environments, yielding drastic improvements in accuracy and detection reliability.
Object frequencies in daily scenes follow a long-tailed distribution. Many objects do not appear frequently enough in scene-centric images (e.g., sightseeing, street views) for us to train accurate object detectors. In contrast, these objects are captured at a higher frequency in object-centric images, which are intended to picture the objects of interest. Motivated by this phenomenon, we propose to take advantage of the object-centric images to improve object detection in scene-centric images. We present a simple yet surprisingly effective framework to do so. On the one hand, our approach turns an object-centric image into a useful training example for object detection in scene-centric images by mitigating the domain gap between the two image sources in both the input and label space. On the other hand, our approach employs a multi-stage procedure to train the object detector, such that the detector learns the diverse object appearances from object-centric images while being tied to the application domain of scene-centric images. On the LVIS dataset, our approach can improve the object detection (and instance segmentation) accuracy of rare objects by 50% (and 33%) relatively, without sacrificing the performance of other classes.
Federated learning aims to leverage users' own data and computational resources in learning a strong global model, without directly accessing their data but only local models. It usually requires multiple rounds of communication, in which aggregating local models into a global model plays an important role. In this paper, we propose a novel aggregation scenario and algorithm named FedDistill, which enjoys the robustness of Bayesian model ensemble in aggregating users' predictions and employs knowledge distillation to summarize the ensemble predictions into a global model, with the help of unlabeled data collected at the server. Our empirical studies validate FedDistill's superior performance, especially when users' data are not i.i.d. and the neural networks go deeper. Moreover, FedDistill is compatible with recent efforts in regularizing users' model training, making it an easily applicable module: you only need to replace the aggregation method but leave other parts of your federated learning algorithms intact.
Semi-supervised domain adaptation (SSDA) aims to adapt models from a labeled source domain to a different but related target domain, from which unlabeled data and a small set of labeled data are provided. In this paper we propose a new approach for SSDA, which is to explicitly decompose SSDA into two sub-problems: a semi-supervised learning (SSL) problem in the target domain and an unsupervised domain adaptation (UDA) problem across domains. We show that these two sub-problems yield very different classifiers, which we leverage with our algorithm MixUp Co-training (MiCo). MiCo applies Mixup to bridge the gap between labeled and unlabeled data of each individual model and employs co-training to exchange the expertise between the two classifiers. MiCo needs no adversarial and minmax training, making it easily implementable and stable. MiCo achieves state-of-the-art results on SSDA datasets, outperforming the prior art by a notable 4% margin on DomainNet.
Existing approaches to depth or disparity estimation output a distribution over a set of pre-defined discrete values. This leads to inaccurate results when the true depth or disparity does not match any of these values. The fact that this distribution is usually learned indirectly through a regression loss causes further problems in ambiguous regions around object boundaries. We address these issues using a new neural network architecture that is capable of outputting arbitrary depth values, and a new loss function that is derived from the Wasserstein distance between the true and the predicted distributions. We validate our approach on a variety of tasks, including stereo disparity and depth estimation, and the downstream 3D object detection. Our approach drastically reduces the error in ambiguous regions, especially around object boundaries that greatly affect the localization of objects in 3D, achieving the state-of-the-art in 3D object detection for autonomous driving.
In the domain of autonomous driving, deep learning has substantially improved the 3D object detection accuracy for LiDAR and stereo camera data alike. While deep networks are great at generalization, they are also notorious to over-fit to all kinds of spurious artifacts, such as brightness, car sizes and models, that may appear consistently throughout the data. In fact, most datasets for autonomous driving are collected within a narrow subset of cities within one country, typically under similar weather conditions. In this paper we consider the task of adapting 3D object detectors from one dataset to another. We observe that naively, this appears to be a very challenging task, resulting in drastic drops in accuracy levels. We provide extensive experiments to investigate the true adaptation challenges and arrive at a surprising conclusion: the primary adaptation hurdle to overcome are differences in car sizes across geographic areas. A simple correction based on the average car size yields a strong correction of the adaptation gap. Our proposed method is simple and easily incorporated into most 3D object detection frameworks. It provides a first baseline for 3D object detection adaptation across countries, and gives hope that the underlying problem may be more within grasp than one may have hoped to believe. Our code is available at https://github.com/cxy1997/3D_adapt_auto_driving.
Reliable and accurate 3D object detection is a necessity for safe autonomous driving. Although LiDAR sensors can provide accurate 3D point cloud estimates of the environment, they are also prohibitively expensive for many settings. Recently, the introduction of pseudo-LiDAR (PL) has led to a drastic reduction in the accuracy gap between methods based on LiDAR sensors and those based on cheap stereo cameras. PL combines state-of-the-art deep neural networks for 3D depth estimation with those for 3D object detection by converting 2D depth map outputs to 3D point cloud inputs. However, so far these two networks have to be trained separately. In this paper, we introduce a new framework based on differentiable Change of Representation (CoR) modules that allow the entire PL pipeline to be trained end-to-end. The resulting framework is compatible with most state-of-the-art networks for both tasks and in combination with PointRCNN improves over PL consistently across all benchmarks -- yielding the highest entry on the KITTI image-based 3D object detection leaderboard at the time of submission. Our code will be made available at https://github.com/mileyan/pseudo-LiDAR_e2e.
We investigate learning a ConvNet classifier with class-imbalanced data. We found that a ConvNet over-fits significantly to the minor classes that do not have sufficient training instances, even if it is trained using vanilla empirical risk minimization (ERM). We conduct a series of analysis and argue that feature deviation between the training and test instances serves as the main cause. We propose to incorporate class-dependent temperatures (CDT) in learning a ConvNet: CDT forces the minor-class instances to have larger decision values in training, so as to compensate for the effect of feature deviation in testing. We validate our approach on several benchmark datasets and achieve promising results. Our studies further suggest that class-imbalanced data affects traditional machine learning and recent deep learning in very different ways. We hope that our insights can inspire new ways of thinking in resolving class-imbalanced deep learning.
In this work, we consider the problem of searching people in an unconstrained environment, with natural language descriptions. Specifically, we study how to systematically design an algorithm to effectively acquire descriptions from humans. An algorithm is proposed by adapting models, used for visual and language understanding, to search a person of interest (POI) in a principled way, achieving promising results without the need to re-design another complicated model. We then investigate an iterative question-answering (QA) strategy that enable robots to request additional information about the POI's appearance from the user. To this end, we introduce a greedy algorithm to rank questions in terms of their significance, and equip the algorithm with the capability to dynamically adjust the length of human-robot interaction according to model's uncertainty. Our approach is validated not only on benchmark datasets but on a mobile robot, moving in a dynamic and crowded environment.
Recent years have witnessed an abundance of new publications and approaches on meta-learning. This community-wide enthusiasm has sparked great insights but has also created a plethora of seemingly different frameworks, which can be hard to compare and evaluate. In this paper, we aim to provide a principled, unifying framework by revisiting and strengthening the connection between meta-learning and traditional supervised learning. By treating pairs of task-specific data sets and target models as (feature, label) samples, we can reduce many meta-learning algorithms to instances of supervised learning. This view not only unifies meta-learning into an intuitive and practical framework but also allows us to transfer insights from supervised learning directly to improve meta-learning. For example, we obtain a better understanding of generalization properties, and we can readily transfer well-understood techniques, such as model ensemble, pre-training, joint training, data augmentation, and even nearest neighbor based methods. We provide an intuitive analogy of these methods in the context of meta-learning and show that they give rise to significant improvements in model performance on few-shot learning.