Large-scale datasets are essential to modern day deep learning. Advocates argue that understanding these methods requires dataset transparency (e.g. "dataset curation, motivation, composition, collection process, etc..."). However, almost no one has suggested the release of the detailed definitions and visual category examples provided to annotators - information critical to understanding the structure of the annotations present in each dataset. These labels are at the heart of public datasets, yet few datasets include the instructions that were used to generate them. We introduce a new task, Labeling Instruction Generation, to address missing publicly available labeling instructions. In Labeling Instruction Generation, we take a reasonably annotated dataset and: 1) generate a set of examples that are visually representative of each category in the dataset; 2) provide a text label that corresponds to each of the examples. We introduce a framework that requires no model training to solve this task and includes a newly created rapid retrieval system that leverages a large, pre-trained vision and language model. This framework acts as a proxy to human annotators that can help to both generate a final labeling instruction set and evaluate its quality. Our framework generates multiple diverse visual and text representations of dataset categories. The optimized instruction set outperforms our strongest baseline across 5 folds by 7.06 mAP for NuImages and 12.9 mAP for COCO.
Vision-language models (VLMs) discriminatively pre-trained with contrastive image-text matching losses such as $P(\text{match}|\text{text}, \text{image})$ have been criticized for lacking compositional understanding. This means they might output similar scores even if the original caption is rearranged into a different semantic statement. To address this, we propose to use the ${\bf V}$isual ${\bf G}$enerative ${\bf P}$re-${\bf T}$raining Score (${\bf VisualGPTScore}$) of $P(\text{text}|\text{image})$, a $\textit{multimodal generative}$ score that captures the likelihood of a text caption conditioned on an image using an image-conditioned language model. Contrary to the belief that VLMs are mere bag-of-words models, our off-the-shelf VisualGPTScore demonstrates top-tier performance on recently proposed image-text retrieval benchmarks like ARO and Crepe that assess compositional reasoning. Furthermore, we factorize VisualGPTScore into a product of the $\textit{marginal}$ P(text) and the $\textit{Pointwise Mutual Information}$ (PMI). This helps to (a) diagnose datasets with strong language bias, and (b) debias results on other benchmarks like Winoground using an information-theoretic framework. VisualGPTScore provides valuable insights and serves as a strong baseline for future evaluation of visio-linguistic compositionality.
Scene flow estimation is the task of describing the 3D motion field between temporally successive point clouds. State-of-the-art methods use strong priors and test-time optimization techniques, but require on the order of tens of seconds for large-scale point clouds, making them unusable as computer vision primitives for real-time applications such as open world object detection. Feed forward methods are considerably faster, running on the order of tens to hundreds of milliseconds for large-scale point clouds, but require expensive human supervision. To address both limitations, we propose Scene Flow via Distillation, a simple distillation framework that uses a label-free optimization method to produce pseudo-labels to supervise a feed forward model. Our instantiation of this framework, ZeroFlow, produces scene flow estimates in real-time on large-scale point clouds at quality competitive with state-of-the-art methods while using zero human labels. Notably, at test-time ZeroFlow is over 1000$\times$ faster than label-free state-of-the-art optimization-based methods on large-scale point clouds and over 1000$\times$ cheaper to train on unlabeled data compared to the cost of human annotation of that data. To facilitate research reuse, we release our code, trained model weights, and high quality pseudo-labels for the Argoverse 2 and Waymo Open datasets.
The open road poses many challenges to autonomous perception, including poor visibility from extreme weather conditions. Models trained on good-weather datasets frequently fail at detection in these out-of-distribution settings. To aid adversarial robustness in perception, we introduce WEDGE (WEather images by DALL-E GEneration): a synthetic dataset generated with a vision-language generative model via prompting. WEDGE consists of 3360 images in 16 extreme weather conditions manually annotated with 16513 bounding boxes, supporting research in the tasks of weather classification and 2D object detection. We have analyzed WEDGE from research standpoints, verifying its effectiveness for extreme-weather autonomous perception. We establish baseline performance for classification and detection with 53.87% test accuracy and 45.41 mAP. Most importantly, WEDGE can be used to fine-tune state-of-the-art detectors, improving SOTA performance on real-world weather benchmarks (such as DAWN) by 4.48 AP for well-generated classes like trucks. WEDGE has been collected under OpenAI's terms of use and is released for public use under the CC BY-NC-SA 4.0 license. The repository for this work and dataset is available at https://infernolia.github.io/WEDGE.
Building animatable 3D models is challenging due to the need for 3D scans, laborious registration, and manual rigging, which are difficult to scale to arbitrary categories. Recently, differentiable rendering provides a pathway to obtain high-quality 3D models from monocular videos, but these are limited to rigid categories or single instances. We present RAC that builds category 3D models from monocular videos while disentangling variations over instances and motion over time. Three key ideas are introduced to solve this problem: (1) specializing a skeleton to instances via optimization, (2) a method for latent space regularization that encourages shared structure across a category while maintaining instance details, and (3) using 3D background models to disentangle objects from the background. We show that 3D models of humans, cats, and dogs can be learned from 50-100 internet videos.
Multi-modal trajectory forecasting methods commonly evaluate using single-agent metrics (marginal metrics), such as minimum Average Displacement Error (ADE) and Final Displacement Error (FDE), which fail to capture joint performance of multiple interacting agents. Only focusing on marginal metrics can lead to unnatural predictions, such as colliding trajectories or diverging trajectories for people who are clearly walking together as a group. Consequently, methods optimized for marginal metrics lead to overly-optimistic estimations of performance, which is detrimental to progress in trajectory forecasting research. In response to the limitations of marginal metrics, we present the first comprehensive evaluation of state-of-the-art (SOTA) trajectory forecasting methods with respect to multi-agent metrics (joint metrics): JADE, JFDE, and collision rate. We demonstrate the importance of joint metrics as opposed to marginal metrics with quantitative evidence and qualitative examples drawn from the ETH / UCY and Stanford Drone datasets. We introduce a new loss function incorporating joint metrics that, when applied to a SOTA trajectory forecasting method, achieves a 7% improvement in JADE / JFDE on the ETH / UCY datasets with respect to the previous SOTA. Our results also indicate that optimizing for joint metrics naturally leads to an improvement in interaction modeling, as evidenced by a 16% decrease in mean collision rate on the ETH / UCY datasets with respect to the previous SOTA.
We address the task of estimating 6D camera poses from sparse-view image sets (2-8 images). This task is a vital pre-processing stage for nearly all contemporary (neural) reconstruction algorithms but remains challenging given sparse views, especially for objects with visual symmetries and texture-less surfaces. We build on the recent RelPose framework which learns a network that infers distributions over relative rotations over image pairs. We extend this approach in two key ways; first, we use attentional transformer layers to process multiple images jointly, since additional views of an object may resolve ambiguous symmetries in any given image pair (such as the handle of a mug that becomes visible in a third view). Second, we augment this network to also report camera translations by defining an appropriate coordinate system that decouples the ambiguity in rotation estimation from translation prediction. Our final system results in large improvements in 6D pose prediction over prior art on both seen and unseen object categories and also enables pose estimation and 3D reconstruction for in-the-wild objects.
We present SLoMo: a first-of-its-kind framework for transferring skilled motions from casually captured "in the wild" video footage of humans and animals to legged robots. SLoMo works in three stages: 1) synthesize a physically plausible reconstructed key-point trajectory from monocular videos; 2) optimize a dynamically feasible reference trajectory for the robot offline that includes body and foot motion, as well as contact sequences that closely tracks the key points; 3) track the reference trajectory online using a general-purpose model-predictive controller on robot hardware. Traditional motion imitation for legged motor skills often requires expert animators, collaborative demonstrations, and/or expensive motion capture equipment, all of which limits scalability. Instead, SLoMo only relies on easy-to-obtain monocular video footage, readily available in online repositories such as YouTube. It converts videos into motion primitives that can be executed reliably by real-world robots. We demonstrate our approach by transferring the motions of cats, dogs, and humans to example robots including a quadruped (on hardware) and a humanoid (in simulation). To the best knowledge of the authors, this is the first attempt at a general-purpose motion transfer framework that imitates animal and human motions on legged robots directly from casual videos without artificial markers or labels.
We explore the task of embodied view synthesis from monocular videos of deformable scenes. Given a minute-long RGBD video of people interacting with their pets, we render the scene from novel camera trajectories derived from in-scene motion of actors: (1) egocentric cameras that simulate the point of view of a target actor and (2) 3rd-person cameras that follow the actor. Building such a system requires reconstructing the root-body and articulated motion of each actor in the scene, as well as a scene representation that supports free-viewpoint synthesis. Longer videos are more likely to capture the scene from diverse viewpoints (which helps reconstruction) but are also more likely to contain larger motions (which complicates reconstruction). To address these challenges, we present Total-Recon, the first method to photorealistically reconstruct deformable scenes from long monocular RGBD videos. Crucially, to scale to long videos, our method hierarchically decomposes the scene motion into the motion of each object, which itself is decomposed into global root-body motion and local articulations. To quantify such "in-the-wild" reconstruction and view synthesis, we collect ground-truth data from a specialized stereo RGBD capture rig for 11 challenging videos, significantly outperforming prior art. Code, videos, and data can be found at https://andrewsonga.github.io/totalrecon .
Current methods for self-supervised LiDAR scene flow estimation work poorly on real data. A variety of flaws in common evaluation protocols have caused leading approaches to focus on problems that do not exist in real data. We analyze a suite of recent works and find that despite their focus on deep learning, the main challenges of the LiDAR scene flow problem -- removing the dominant rigid motion and robustly estimating the simple motions that remain -- can be more effectively solved with classical techniques such as ICP motion compensation and enforcing piecewise rigid assumptions. We combine these steps with a test-time optimization method to form a state-of-the-art system that does not require any training data. Because our final approach is dataless, it can be applied on different datasets with diverse LiDAR rigs without retraining. Our proposed approach outperforms all existing methods on Argoverse 2.0, halves the error rate on NuScenes, and even rivals the performance of supervised networks on Waymo and lidarKITTI.