Multi-modal large language models (MLLMs) have shown incredible capabilities in a variety of 2D vision and language tasks. We extend MLLMs’ perceptual capabilities to ground and reason about images in 3-dimensional space. To that end, we first develop a large-scale pretraining dataset for 2D and 3D called LV3D by combining multiple existing 2D and 3D recognition datasets under a common task formulation: as multi-turn question-answering. Next, we introduce a new MLLM named CUBE-LLM and pre-train it on LV3D. We show that pure data scaling makes a strong 3D perception capability without 3D specific architectural design or training objective. CUBE-LLM exhibits intriguing properties similar to LLMs: (1) CUBE-LLM can apply chain-of-thought prompting to improve 3D understanding from 2D context information. (2) CUBE-LLM can follow complex and diverse instructions and adapt to versatile input and output formats. (3) CUBE-LLM can be visually prompted such as 2D box or a set of candidate 3D boxes from specialists. Our experiments on outdoor benchmarks demonstrate that CUBE-LLM significantly outperforms existing baselines by 21.3 points of AP-BEV on the Talk2Car dataset for 3D grounded reasoning and 17.7 points on the DriveLM dataset for complex reasoning about driving scenarios, respectively. CUBE-LLM also shows competitive results in general MLLM benchmarks such as refCOCO for 2D grounding with (87.0) average score, as well as visual question answering benchmarks such as VQAv2, GQA, SQA, POPE, etc. for complex reasoning.

Autonomous driving necessitates the ability to reason about future interactions between traffic agents and to make informed evaluations for planning. This paper introduces the Gen-Drive framework, which shifts from the traditional prediction and deterministic planning framework to a generation-then-evaluation planning paradigm. The framework employs a behavior diffusion model as a scene generator to produce diverse possible future scenarios, thereby enhancing the capability for joint interaction reasoning. To facilitate decision-making, we propose a scene evaluator (reward) model, trained with pairwise preference data collected through VLM assistance, thereby reducing human workload and enhancing scalability. Furthermore, we utilize an RL fine-tuning framework to improve the generation quality of the diffusion model, rendering it more effective for planning tasks. We conduct training and closed-loop planning tests on the nuPlan dataset, and the results demonstrate that employing such a generation-then-evaluation strategy outperforms other learning-based approaches. Additionally, the fine-tuned generative driving policy shows significant enhancements in planning performance. We further demonstrate that utilizing our learned reward model for evaluation or RL fine-tuning leads to better planning performance compared to relying on human-designed rewards. Project website: https://mczhi.github.io/GenDrive.

Recent works have proposed end-to-end autonomous vehicle (AV) architectures comprised of differentiable modules, achieving state-of-the-art driving performance. While they provide advantages over the traditional perception-prediction-planning pipeline (e.g., removing information bottlenecks between components and alleviating integration challenges), they do so using a diverse combination of tasks, modules, and their interconnectivity. As of yet, however, there has been no systematic analysis of the necessity of these modules or the impact of their connectivity, placement, and internal representations on overall driving performance. Addressing this gap, our work conducts a comprehensive exploration of the design space of end-to-end modular AV stacks. Our findings culminate in the development of PARA-Drivel: a fully parallel end-to-end AV architecture. PARA-Drive not only achieves state-of-the-art performance in perception, prediction, and planning, but also significantly enhances runtime speed by nearly 3 x, without compromising on interpretability or safety.

We present EmerNeRF, a simple yet powerful approach for learning spatial-temporal representations of dynamic driving scenes. Grounded in neural fields, EmerNeRF simultaneously captures scene geometry, appearance, motion, and semantics via self-bootstrapping. EmerNeRF hinges upon two core components: First, it stratifies scenes into static and dynamic fields. This decomposition emerges purely from self-supervision, enabling our model to learn from general, in-the-wild data sources. Second, EmerNeRF parameterizes an induced flow field from the dynamic field and uses this flow field to further aggregate multi-frame features, amplifying the rendering precision of dynamic objects. Coupling these three fields (static, dynamic, and flow) enables EmerNeRF to represent highly-dynamic scenes self-sufficiently, without relying on ground truth object annotations or pre-trained models for dynamic object segmentation or optical flow estimation. Our method achieves state-of-the-art performance in sensor simulation, significantly outperforming previous methods when reconstructing static (+2.93 PSNR) and dynamic (+3.70 PSNR) scenes. In addition, to bolster EmerNeRF's semantic generalization, we lift 2D visual foundation model features into 4D space-time and address a general positional bias in modern Transformers, significantly boosting 3D perception performance (e.g., 37.50% relative improvement in occupancy prediction accuracy on average). Finally, we construct a diverse and challenging 120-sequence dataset to benchmark neural fields under extreme and highly-dynamic settings. See the project page for code, data, and request pre-trained models: https://emernerf.github.io

Autonomous driving has traditionally relied heavily on costly and labor-intensive High Definition (HD) maps, hindering scalability. In contrast, Standard Definition (SD) maps are more affordable and have worldwide coverage, offering a scalable alternative. In this work, we systematically explore the effect of SD maps for real-time lane-topology understanding. We propose a novel framework to integrate SD maps into online map prediction and propose a Transformer-based encoder, SD Map Encoder Representations from transFormers, to leverage priors in SD maps for the lane-topology prediction task. This enhancement consistently and significantly boosts (by up to 60%) lane detection and topology prediction on current state-of-the-art online map prediction methods without bells and whistles and can be immediately incorporated into any Transformer-based lane-topology method. Code is available at https://github.com/NVlabs/SMERF.

Benchmarking vision-based driving policies is challenging. On one hand, open-loop evaluation with real data is easy, but these results do not reflect closed-loop performance. On the other, closed-loop evaluation is possible in simulation, but is hard to scale due to its significant computational demands. Further, the simulators available today exhibit a large domain gap to real data. This has resulted in an inability to draw clear conclusions from the rapidly growing body of research on end-to-end autonomous driving. In this paper, we present NAVSIM, a middle ground between these evaluation paradigms, where we use large datasets in combination with a non-reactive simulator to enable large-scale real-world benchmarking. Specifically, we gather simulation-based metrics, such as progress and time to collision, by unrolling bird's eye view abstractions of the test scenes for a short simulation horizon. Our simulation is non-reactive, i.e., the evaluated policy and environment do not influence each other. As we demonstrate empirically, this decoupling allows open-loop metric computation while being better aligned with closed-loop evaluations than traditional displacement errors. NAVSIM enabled a new competition held at CVPR 2024, where 143 teams submitted 463 entries, resulting in several new insights. On a large set of challenging scenarios, we observe that simple methods with moderate compute requirements such as TransFuser can match recent large-scale end-to-end driving architectures such as UniAD. Our modular framework can potentially be extended with new datasets, data curation strategies, and metrics, and will be continually maintained to host future challenges. Our code is available at https://github.com/autonomousvision/navsim.

Kalman filter (KF) based methods for multi-object tracking (MOT) make an assumption that objects move linearly. While this assumption is acceptable for very short periods of occlusion, linear estimates of motion for prolonged time can be highly inaccurate. Moreover, when there is no measurement available to update Kalman filter parameters, the standard convention is to trust the priori state estimations for posteriori update. This leads to the accumulation of errors during a period of occlusion. The error causes significant motion direction variance in practice. In this work, we show that a basic Kalman filter can still obtain state-of-the-art tracking performance if proper care is taken to fix the noise accumulated during occlusion. Instead of relying only on the linear state estimate (i.e., estimation-centric approach), we use object observations (i.e., the measurements by object detector) to compute a virtual trajectory over the occlusion period to fix the error accumulation of filter parameters. This allows more time steps to correct errors accumulated during occlusion. We name our method Observation-Centric SORT (OC-SORT). It remains Simple, Online, and Real-Time but improves robustness during occlusion and non-linear motion. Given off-the-shelf detections as input, OC-SORT runs at 700+ FPS on a single CPU. It achieves state-of-the-art on multiple datasets, including MOT17, MOT20, KITTI, head tracking, and especially DanceTrack where the object motion is highly non-linear. The code and models are available at https://github.com/noahcao/OC_SORT.

Autonomous vehicles (AVs) need to reason about the multimodal behavior of neighboring agents while planning their own motion. Many existing trajectory planners seek a single trajectory that performs well under all plausible futures simultaneously, ignoring bi-directional interactions and thus leading to overly conservative plans. Policy planning, whereby the ego agent plans a policy that reacts to the environment's multimodal behavior, is a promising direction as it can account for the action-reaction interactions between the AV and the environment. However, most existing policy planners do not scale to the complexity of real autonomous vehicle applications: they are either not compatible with modern deep learning prediction models, not interpretable, or not able to generate high quality trajectories. To fill this gap, we propose Tree Policy Planning (TPP), a policy planner that is compatible with state-of-the-art deep learning prediction models, generates multistage motion plans, and accounts for the influence of ego agent on the environment behavior. The key idea of TPP is to reduce the continuous optimization problem into a tractable discrete Markov Decision Process (MDP) through the construction of two tree structures: an ego trajectory tree for ego trajectory options, and a scenario tree for multi-modal ego-conditioned environment predictions. We demonstrate the efficacy of TPP in closed-loop simulations based on real-world nuScenes dataset and results show that TPP scales to realistic AV scenarios and significantly outperforms non-policy baselines.

Multi-agent trajectory prediction is critical for planning and decision-making in human-interactive autonomous systems, such as self-driving cars. However, most prediction models are developed separately from their upstream perception (detection and tracking) modules, assuming ground truth past trajectories as inputs. As a result, their performance degrades significantly when using real-world noisy tracking results as inputs. This is typically caused by the propagation of errors from tracking to prediction, such as noisy tracks, fragments and identity switches. To alleviate this propagation of errors, we propose a new prediction paradigm that uses detections and their affinity matrices across frames as inputs, removing the need for error- prone data association during tracking. Since affinity matrices contain “soft” information about the similarity and identity of detections across frames, making prediction directly from affinity matrices retains strictly more information than making prediction from the tracklets generated by data association. Experiments on large-scale, real-world autonomous driving datasets show that our affinity-based prediction scheme11Our project website is at https://www.xinshuoweng.com/projects/Affinipred. reduces overall prediction errors by up to 57.9%, in comparison to standard prediction pipelines that use tracklets as inputs, with even more significant error reduction (up to 88.6%) if restricting the evaluation to challenging scenarios with tracking errors.

There has been tremendous progress in the development of individual modules of the standard perception-prediction-planning robot autonomy stack. However, the principled integration of these modules has received less attention, particularly in terms of cascading errors. In this work, we both characterize and address the problem of cascading errors, focusing on the coupling between tracking and prediction. First, we comprehensively evaluate the impact of tracking errors on prediction performance with modern tracking and prediction methods on real-world data. We find that prediction methods experience a significant (even order of magnitude) drop in performance when consuming tracked trajectories as inputs (typical in practice), compared to the idealized setting where ground truth past trajectories are used as inputs. To address this issue, we propose a multi-hypothesis tracking and prediction framework. Rather than relying on a single set of tracking results for prediction, we simultaneously reason about multiple sets of tracking results, thereby increasing the likelihood of including accurate tracking results as inputs to prediction. We show that our framework1improves overall prediction performance over the standard single-hypothesis tracking-prediction pipeline by up to 34.2% on the nuScenes dataset, with even more significant improvements (up to $\sim 70$%) when restricting evaluation to challenging scenarios involving identity switches and fragments, all with an acceptable computation overhead.

We address the problem of estimating the 3D pose of a network of cameras for large-environment wide-baseline scenarios, e.g., cameras for construction sites, sports stadiums, and public spaces. This task is challenging since detecting and matching the same 3D keypoint observed from two very different camera views is difficult, making standard structure-from-motion (SfM) pipelines inapplicable. In such circumstances, treating people in the scene as "keypoints" and associating them across different camera views can be an alternative method for obtaining correspondences. Based on this intuition, we propose a method that uses ideas from person re-identification (re-ID) for wide-baseline camera calibration. Our method first employs a re-ID method to associate human bounding boxes across cameras, then converts bounding box correspondences to point correspondences, and finally solves for camera pose using multi-view geometry and bundle adjustment. Since our method does not require specialized calibration targets except for visible people, it applies to situations where frequent calibration updates are required. We perform extensive experiments on datasets captured from scenes of different sizes (80m2, 350m2, 600m2), camera settings (indoor and outdoor), and human activities (walking, playing basketball, construction). Experiment results show that our method achieves similar performance to standard SfM methods relying on manually labeled point correspondences.

Predicting accurate future trajectories of multiple agents is essential for autonomous systems but is challenging due to the complex interaction between agents and the uncertainty in each agent’s future behavior. Forecasting multi-agent trajectories requires modeling two key dimensions: (1) time dimension, where we model the influence of past agent states over future states; (2) social dimension, where we model how the state of each agent affects others. Most prior methods model these two dimensions separately, e.g., first using a temporal model to summarize features over time for each agent independently and then modeling the interaction of the summarized features with a social model. This approach is suboptimal since independent feature encoding over either the time or social dimension can result in a loss of information. Instead, we would prefer a method that allows an agent’s state at one time to directly affect another agent’s state at a future time. To this end, we propose a new Transformer, termed AgentFormer, that simultaneously models the time and social dimensions. The model leverages a sequence representation of multi-agent trajectories by flattening trajectory features across time and agents. Since standard attention operations disregard the agent identity of each element in the sequence, AgentFormer uses a novel agent-aware attention mechanism that preserves agent identities by attending to elements of the same agent differently than elements of other agents. Based on AgentFormer, we propose a stochastic multi-agent trajectory prediction model that can attend to features of any agent at any previous timestep when inferring an agent’s future position. The latent intent of all agents is also jointly modeled, allowing the stochasticity in one agent’s behavior to affect other agents. Extensive experiments show that our method significantly improves the state of the art on well-established pedestrian and autonomous driving datasets.

We address the task of Re-Identification (Re-ID) in multi-target multi-camera (MTMC) tracking where we track multiple pedestrians using multiple overlapping uncalibrated (unknown pose) cameras. Since the videos are temporally synchronized and spatially overlapping, we can see a person from multiple views and associate their trajectory across cameras. In order to find the correct association between pedestrians visible from multiple views during the same time window, we extract a visual feature from a tracklet (sequence of pedestrian images) that encodes its similarity and dissimilarity to all other candidate tracklets. We propose a inter-tracklet (person to person) attention mechanism that learns a representation for a target tracklet while taking into account other tracklets across multiple views. Furthermore, to encode the gait and motion of a person, we introduce second intra-tracklet (person-specific) attention module with position embeddings. This second module employs a transformer encoder to learn a feature from a sequence of features over one tracklet. Experimental results on WILDTRACK and our new dataset ‘ConstructSite’ confirm the superiority of our model over state-of-the-art ReID methods (5% and 10% performance gain respectively) in the context of uncalibrated MTMC tracking. While our model is designed for overlapping cameras, we also obtain state-of-the-art results on two other benchmark datasets (MARS and DukeMTMC) with non-overlapping cameras.

LiDAR sensors can be used to obtain a wide range of measurement signals other than a simple 3D point cloud, and those signals can be leveraged to improve perception tasks like 3D object detection. A single laser pulse can be partially reflected by multiple objects along its path, resulting in multiple measurements called echoes. Multi-echo measurement can provide information about object contours and semi-transparent surfaces which can be used to better identify and locate objects. LiDAR can also measure surface reflectance (intensity of laser pulse return), as well as ambient light of the scene (sunlight reflected by objects). These signals are already available in commercial LiDAR devices but have not been used in most LiDAR-based detection models. We present a 3D object detection model which leverages the full spectrum of measurement signals provided by LiDAR. First, we propose a multi-signal fusion (MSF) module to combine (1) the reflectance and ambient features extracted with a 2D CNN, and (2) point cloud features extracted using a 3D graph neural network (GNN). Second, we propose a multi-echo aggregation (MEA) module to combine the information encoded in different sets of echo points. Compared with traditional single echo point cloud methods, our proposed Multi-Signal LiDAR Detector (MSLiD) extracts richer context information from a wider range of sensing measurements and achieves more accurate 3D object detection. Experiments show that by incorporating the multi-modality of LiDAR, our method outperforms the state-of-the-art by up to relatively 9.1%.

Object detection and data association are critical components in multi-object tracking (MOT) systems. Despite the fact that the two components are dependent on each other, prior works often design detection and data association modules separately which are trained with separate objectives. As a result, one cannot back-propagate the gradients and optimize the entire MOT system, which leads to sub-optimal performance. To address this issue, recent works simultaneously optimize detection and data association modules under a joint MOT framework, which has shown improved performance in both modules. In this work, we propose a new instance of joint MOT approach based on Graph Neural Networks (GNNs). The key idea is that GNNs can model relations between variablesized objects in both the spatial and temporal domains, which is essential for learning discriminative features for detection and data association. Through extensive experiments on the MOT15/16/17/20 datasets, we demonstrate the effectiveness of our GNN-based joint MOT approach and show state-of-the-art performance for both detection and MOT tasks.

Person re-identification (re-ID) aims to recognize instances of the same person contained in multiple images taken across different cameras. Existing methods for re-ID tend to rely heavily on the assumption that both query and gallery images of the same person have the same clothing. Unfortunately, this assumption may not hold for datasets captured over long periods of time. To tackle the re-ID problem in the context of clothing changes, we propose a novel representation learning method which is able to generate a shape-based feature representation that is invariant to clothing. We call our model the Clothing Agnostic Shape Extraction Network (CASE-Net). CASE-Net learns a representation of a person that depends primarily on shape via adversarial learning and feature disentanglement. Quantitative and qualitative results across 5 datasets (Div-Market, Market1501, three large-scale datasets under clothing changes) show our approach makes significant improvements over prior state-of-the-art approaches.

We present supervision by registration and triangulation (SRT), an unsupervised approach that utilizes unlabeled multi-view video to improve the accuracy and precision of landmark detectors. Being able to utilize unlabeled data enables our detectors to learn from massive amounts of unlabeled data freely available and not be limited by the quality and quantity of manual human annotations. To utilize unlabeled data, there are two key observations: (I) The detections of the same landmark in adjacent frames should be coherent with registration, i.e., optical flow. (II) The detections of the same landmark in multiple synchronized and geometrically calibrated views should correspond to a single 3D point, i.e., multi-view consistency. Registration and multi-view consistency are sources of supervision that do not require manual labeling, thus it can be leveraged to augment existing training data during detector training. End-to-end training is made possible by differentiable registration and 3D triangulation modules. Experiments with 11 datasets and a newly proposed metric to measure precision demonstrate accuracy and precision improvements in landmark detection on both images and video.

3D Multi-object tracking (MOT) is crucial to autonomous systems. Recent work uses a standard tracking-by-detection pipeline, where feature extraction is first performed independently for each object in order to compute an affinity matrix. Then the affinity matrix is passed to the Hungarian algorithm for data association. A key process of this standard pipeline is to learn discriminative features for different objects in order to reduce confusion during data association. In this work, we propose two techniques to improve the discriminative feature learning for MOT: (1) instead of obtaining features for each object independently, we propose a novel feature interaction mechanism by introducing the Graph Neural Network. As a result, the feature of one object is informed of the features of other objects so that the object feature can lean towards the object with similar feature (i.e., object probably with a same ID) and deviate from objects with dissimilar features (i.e., object probably with different IDs), leading to a more discriminative feature for each object; (2) instead of obtaining the feature from either 2D or 3D space in prior work, we propose a novel joint feature extractor to learn appearance and motion features from 2D and 3D space simultaneously. As features from different modalities often have complementary information, the joint feature can be more discriminate than feature from each individual modality. To ensure that the joint feature extractor does not heavily rely on one modality, we also propose an ensemble training paradigm. Through extensive evaluation, our proposed method achieves state-of-the-art performance on KITTI and nuScenes 3D MOT benchmarks. Our code will be made available at https://github.com/xinshuoweng/GNN3DMOT.

We aim to enable robots to visually localize a target person through the aid of an additional sensing modality - the target person's 3D inertial measurements. The need for such technology may arise when a robot is to meet a person in a crowd for the first time or when an autonomous vehicle must rendezvous with a rider amongst a crowd without knowing the appearance of the person in advance. A person's inertial information can be measured with a wearable device such as a smart-phone and can be shared selectively with an autonomous system during the rendezvous. We propose a method to learn a visual-inertial feature space in which the motion of a person in video can be easily matched to the motion measured by a wearable inertial measurement unit (IMU). The transformation of the two modalities into the joint feature space is learned through the use of a triplet loss which forces inertial motion features and video motion features generated by the same person to lie close in the joint feature space. To validate our approach, we compose a dataset of over 3,000 video segments of moving people along with wearable IMU data. We show that our method is able to localize a target person with 80.7% accuracy averaged over testing data with various number of candidates using only 5 seconds of IMU data and video.

3D multi-object tracking (MOT) is an essential component for many applications such as autonomous driving and assistive robotics. Recent work on 3D MOT focuses on developing accurate systems giving less attention to practical considerations such as computational cost and system complexity. In contrast, this work proposes a simple real-time 3D MOT system. Our system first obtains 3D detections from a LiDAR point cloud. Then, a straightforward combination of a 3D Kalman filter and the Hungarian algorithm is used for state estimation and data association. Additionally, 3D MOT datasets such as KITTI evaluate MOT methods in the 2D space and standardized 3D MOT evaluation tools are missing for a fair comparison of 3D MOT methods. Therefore, we propose a new 3D MOT evaluation tool along with three new metrics to comprehensively evaluate 3D MOT methods. We show that, although our system employs a combination of classical MOT modules, we achieve state-of-the-art 3D MOT performance on two 3D MOT benchmarks (KITTI and nuScenes). Surprisingly, although our system does not use any 2D data as inputs, we achieve competitive performance on the KITTI 2D MOT leaderboard. Our proposed system runs at a rate of 207.4 FPS on the KITTI dataset, achieving the fastest speed among all modern MOT systems. To encourage standardized 3D MOT evaluation, our code is publicly available at http://www.xinshuoweng.com/projects/AB3DMOT.

We explore the possibility of using a single monocular camera to forecast the time to collision between a suitcase-shaped robot being pushed by its user and other nearby pedestrians. We develop a purely image-based deep learning approach that directly estimates the time to collision without the need of relying on explicit geometric depth estimates or velocity information to predict future collisions. While previous work has focused on detecting immediate collision in the context of navigating Unmanned Aerial Vehicles, the detection was limited to a binary variable (i.e., collision or no collision). We propose a more fine-grained approach to collision forecasting by predicting the exact time to collision in terms of milliseconds, which is more helpful for collision avoidance in the context of dynamic path planning. To evaluate our method, we have collected a novel dataset of over 13,000 indoor video segments each showing a trajectory of at least one person ending in a close proximity (a near collision) with the camera mounted on a mobile suitcase-shaped platform. Using this dataset, we do extensive experimentation on different temporal windows as input using an exhaustive list of state-of-the-art convolutional neural networks (CNNs). Our results show that our proposed multi-stream CNN is the best model for predicting time to near-collision. The average prediction error of our time to near-collision is 0.75 seconds across the test videos. The project webpage can be found at https://aashi7.github.io/NearCollision.html.

We focus on estimating the 3D orientation of the ground plane from a single image. We formulate the problem as an inter-mingled multi-task prediction problem by jointly optimizing for pixel-wise surface normal direction, ground plane segmentation, and depth estimates. Specifically, our proposed model, GroundNet, first estimates the depth and surface normal in two separate streams, from which two ground plane normals are then computed deterministically. To leverage the geometric correlation between depth and normal, we propose to add a consistency loss on top of the computed ground plane normals. In addition, a ground segmentation stream is used to isolate the ground regions so that we can selectively back-propagate parameter updates through only the ground regions in the image. Our method achieves the top-ranked performance on ground plane normal estimation and horizon line detection on the real-world outdoor datasets of ApolloScape and KITTI, improving the performance of previous art by up to 17.7% relatively.

In this paper, we present supervision-by-registration, an unsupervised approach to improve the precision of facial landmark detectors on both images and video. Our key observation is that the detections of the same landmark in adjacent frames should be coherent with registration, i.e., optical flow. Interestingly, coherency of optical flow is a source of supervision that does not require manual labeling, and can be leveraged during detector training. For example, we can enforce in the training loss function that a detected landmark at framet-1followed by optical flow tracking from framet-1to frametshould coincide with the location of the detection at framet. Essentially, supervision-by-registration augments the training loss function with a registration loss, thus training the detector to have output that is not only close to the annotations in labeled images, but also consistent with registration on large amounts of unlabeled videos. End-to-end training with the registration loss is made possible by a differentiable Lucas-Kanade operation, which computes optical flow registration in the forward pass, and back-propagates gradients that encourage temporal coherency in the detector. The output of our method is a more precise image-based facial landmark detector, which can be applied to single images or video. With supervision-by-registration, we demonstrate (1) improvements in facial landmark detection on both images (300W, ALFW) and video (300VW, Youtube-Celebrities), and (2) significant reduction of jittering in video detections.

Across most pedestrian detection datasets, it is typically assumed that pedestrians will be standing upright with respect to the image coordinate system. This assumption is not always valid for many vision-equipped mobile platforms, such as mobile phones, UAVs, or construction vehicles on rugged terrain. In these situations, the motion of the camera can cause images of pedestrians to be captured at extreme angles. This can lead to inferior pedestrian detection performance when using standard pedestrian detectors. To address this issue, we propose a Rotational Rectification Network (R2N) that can be inserted into any CNNbased pedestrian (or object) detector to adapt it to significant changes in camera rotation. The rotational rectification network uses a 2D rotation estimation module that passes rotational information to a spatial transformer network [12] to undistort image features. To enable robust rotation estimation, we propose a Global Polar Pooling (GPPooling) operator to capture rotational shifts in convolutional features. Through our experiments, we show how our rotational rectification network can be used to improve the performance of state-of-the-art pedestrian detectors under heavy image rotation by up to 45%.