With the Autonomous Vehicle (AV) industry shifting towards machine-learned approaches for motion plan-ning [1], the performance of self-driving systems is starting to rely heavily on large quantities of expert driving demon-strations. However, collecting this demonstration data typically involves expensive HD sensor suites (LiDAR + RADAR + cameras), which quickly becomes financially infeasible at the scales required. This motivates the use of commodity sensors like cameras for data collection, which are an order of mag-nitude cheaper than HD sensor suites, but offer lower fidelity. Leveraging these sensors for training an AV motion planner opens a financially viable path to observe the ‘long tail’ of driving events. As our main contribution we show it is possible to train a high-performance motion planner using commodity vision data which outperforms planners trained on HD-sensor data for a fraction of the cost. To the best of our knowledge, we are the first to demonstrate this using real-world data. We compare the performance of the autonomy system on these two different sensor configurations, and show that we can compensate for the lower sensor fidelity by means of increased quantity: a planner trained on 100h of commodity vision data outperforms the one with 25h of expensive HD data (see Fig. 1). We also share the engineering challenges we had to tackle to make this work.

In this work we present a simple end-to-end trainable machine learning system capable of realistically simulating driving experiences. This can be used for verification of self-driving system performance without relying on expensive and time-consuming road testing. In particular, we frame the simulation problem as a Markov Process, leveraging deep neural networks to model both state distribution and transition function. These are trainable directly from the existing raw observations without the need of any handcrafting in the form of plant or kinematic models. All that is needed is a dataset of historical traffic episodes. Our formulation allows the system to construct never seen scenes that unfold realistically reacting to the self-driving car’s behaviour. We train our system directly from 1,000 hours of driving logs and measure both realism, reactivity of the simulation as the two key properties of the simulation. At the same time we apply the method to evaluate performance of a recently proposed state-of-the-art ML planning system [1] trained from human driving logs. We discover this planning system is prone to previously unreported causal confusion issues that are difficult to test by non-reactive simulation. To the best of our knowledge, this is the first work that directly merges highly realistic data-driven simulations with a closed loop evaluation for self-driving vehicles. We make the data, code, and pre-trained models publicly available to further stimulate simulation development.

We investigate what grade of sensor data is required for training an imitation-learning-based AV planner on human expert demonstration. Machine-learned planners [1] are very hungry for training data, which is usually collected using vehicles equipped with the same sensors used for autonomous operation [1]. This is costly and non-scalable. If cheaper sensors could be used for collection instead, data availability would go up, which is crucial in a field where data volume requirements are large and availability is small. We present experiments using up to 1000 hours worth of expert demonstration and find that training with 10x lower-quality data outperforms 1x AV-grade data in terms of planner performance (see Fig. 1). The important implication of this is that cheaper sensors can indeed be used. This serves to improve data access and democratize the field of imitation-based motion planning. Alongside this, we perform a sensitivity analysis of planner performance as a function of perception range, field-of-view, accuracy, and data volume, and reason about why lower-quality data still provide good planning results.

In this paper we present the first published end-to-end production computer-vision system for powering city-scale shared augmented reality experiences on mobile devices. In doing so we propose a new formulation for an experience-based mapping framework as an effective solution to the key issues of city-scale SLAM scalability, robustness, map updates and all-time all-weather performance required by a production system. Furthermore, we propose an effective way of synchronising SLAM systems to deliver seamless real-time localisation of multiple edge devices at the same time. All this in the presence of network latency and bandwidth limitations. The resulting system is deployed and tested at scale in San Francisco where it delivers AR experiences in a mapped area of several hundred kilometers. To foster further development of this area we offer the data set to the public, constituting the largest of this kind to date.

We propose an automatic system for organizing the content of a collection of unstructured videos of an articulated object class (e.g., tiger, horse). By exploiting the recurring motion patterns of the class across videos, our system: (1) identifies its characteristic behaviors, and (2) recovers pixel-to-pixel alignments across different instances. Our system can be useful for organizing video collections for indexing and retrieval. Moreover, it can be a platform for learning the appearance or behaviors of object classes from Internet video. Traditional supervised techniques cannot exploit this wealth of data directly, as they require a large amount of time-consuming manual annotations. The behavior discovery stage generates temporal video intervals, each automatically trimmed to one instance of the discovered behavior, clustered by type. It relies on our novel motion representation for articulated motion based on the displacement of ordered pairs of trajectories. The alignment stage aligns hundreds of instances of the class to a great accuracy despite considerable appearance variations (e.g., an adult tiger and a cub). It uses a flexible thin plate spline deformation model that can vary through time. We carefully evaluate each step of our system on a new, fully annotated dataset. On behavior discovery, we outperform the state-of-the-art improved dense trajectory feature descriptor. On spatial alignment, we outperform the popular SIFT Flow algorithm.

We propose a motion-based method to discover the physical parts of an articulated object class (e.g. head/torso/leg of a horse) from multiple videos. The key is to find object regions that exhibit consistent motion relative to the rest of the object, across multiple videos. We can then learn a location model for the parts and segment them accurately in the individual videos using an energy function that also enforces temporal and spatial consistency in part motion. Unlike our approach, traditional methods for motion segmentation or non-rigid structure from motion operate on one video at a time. Hence they cannot discover a part unless it displays independent motion in that particular video. We evaluate our method on a new dataset of 32 videos of tigers and horses, where we significantly outperform a recent motion segmentation method on the task of part discovery (obtaining roughly twice the accuracy).

We propose an unsupervised approach for discovering characteristic motion patterns in videos of highly articulated objects performing natural, unscripted behaviors, such as tigers in the wild. We discover consistent patterns in a bottom-up manner by analyzing the relative displacements of large numbers of ordered trajectory pairs through time, such that each trajectory is attached to a different moving part on the object. The pairs of trajectories descriptor relies entirely on motion and is more discriminative than state-of-the-art features that employ single trajectories. Our method generates temporal video intervals, each automatically trimmed to one instance of the discovered behavior, and clusters them by type (e.g., running, turning head, drinking water). We present experiments on two datasets: dogs from YouTube-Objects and a new dataset of National Geographic tiger videos. Results confirm that our proposed descriptor outperforms existing appearance- and trajectory-based descriptors (e.g., HOG and DTFs) on both datasets and enables us to segment unconstrained animal video into intervals containing single behaviors.

We develop a comprehensive Bayesian generative model for understanding indoor scenes. While it is common in this domain to approximate objects with 3D bounding boxes, we propose using strong representations with finer granularity. For example, we model a chair as a set of four legs, a seat and a backrest. We find that modeling detailed geometry improves recognition and reconstruction, and enables more refined use of appearance for scene understanding. We demonstrate this with a new likelihood function that rewards 3D object hypotheses whose 2D projection is more uniform in color distribution. Such a measure would be confused by background pixels if we used a bounding box to represent a concave object like a chair. Complex objects are modeled using a set or re-usable 3D parts, and we show that this representation captures much of the variation among object instances with relatively few parameters. We also designed specific data-driven inference mechanisms for each part that are shared by all objects containing that part, which helps make inference transparent to the modeler. Further, we show how to exploit contextual relationships to detect more objects, by, for example, proposing chairs around and underneath tables. We present results showing the benefits of each of these innovations. The performance of our approach often exceeds that of state-of-the-art methods on the two tasks of room layout estimation and object recognition, as evaluated on two bench mark data sets used in this domain.

We develop a Bayesian modeling approach for tracking people in 3D from monocular video with unknown cameras. Modeling in 3D provides natural explanations for occlusions and smoothness discontinuities that result from projection, and allows priors on velocity and smoothness to be grounded in physical quantities: meters and seconds vs. pixels and frames. We pose the problem in the context of data association, in which observations are assigned to tracks. A correct application of Bayesian inference to multi-target tracking must address the fact that the model's dimension changes as tracks are added or removed, and thus, posterior densities of different hypotheses are not comparable. We address this by marginalizing out the trajectory parameters so the resulting posterior over data associations has constant dimension. This is made tractable by using (a) Gaussian process priors for smooth trajectories and (b) approximately Gaussian likelihood functions. Our approach provides a principled method for incorporating multiple sources of evidence, we present results using both optical flow and object detector outputs. Results are comparable to recent work on 3D tracking and, unlike others, our method requires no pre-calibrated cameras.

We propose a method for understanding the 3D geometry of indoor environments (e.g. bedrooms, kitchens) while simultaneously identifying objects in the scene (e.g. beds, couches, doors). We focus on how modeling the geometry and location of specific objects is helpful for indoor scene understanding. For example, beds are shorter than they are wide, and are more likely to be in the center of the room than cabinets, which are tall and narrow. We use a generative statistical model that integrates a camera model, an enclosing room “box”, frames (windows, doors, pictures), and objects (beds, tables, couches, cabinets), each with their own prior on size, relative dimensions, and locations. We fit the parameters of this complex, multi-dimensional statistical model using an MCMC sampling approach that combines discrete changes (e.g, adding a bed), and continuous parameter changes (e.g., making the bed larger). We find that introducing object category leads to state-of-the-art performance on room layout estimation, while also enabling recognition based only on geometry.

We propose a top down approach for understanding indoor scenes such as bedrooms and living rooms. These environments typically have the Manhattan world property that many surfaces are parallel to three principle ones. Further, the 3D geometry of the room and objects within it can largely be approximated by non overlapping simple structures such as single blocks (e.g. the room boundary), thin blocks (e.g. picture frames), and objects that are well modeled by single blocks (e.g. simple beds). We separately model the 3D geometry, the imaging process (camera parameters), and edge likelihood, to provide a generative statistical model for image data. We fit this model using data driven MCMC sampling. We combine reversible jump Metropolis Hastings samples for discrete changes in the model such as the number of blocks, and stochastic dynamics to estimate continuous parameter values in a particular parameter space that includes block positions, block sizes, and camera parameters. We tested our approach on two datasets using room box pixel orientation. Despite using only bounding box geometry and, in particular, not training on appearance, our method achieves results approaching those of others. We also introduce a new evaluation method for this domain based on ground truth camera parameters, which we found to be more sensitive to the task of understanding scene geometry.

We present a method for automatically aligning words to image regions that integrates specific object classifiers (e.g., "car" detectors) with weak models based on appearance features. Previous strategies have largely focused on the latter, and thus have not exploited progress on object category recognition. Hence, we augment region labeling with object detection, which simplifies the problem by reliably identifying a subset of the labels, and thereby reducing correspondence ambiguity overall. Comprehensive testing on the SAIAPR TC dataset shows that principled integration of object detection improves the region labeling task.