Yahoo! Labs
Abstract:Recent advancements in Vision-Language Models (VLMs) have demonstrated strong potential for autonomous driving tasks. However, their spatial understanding and reasoning-key capabilities for autonomous driving-still exhibit significant limitations. Notably, none of the existing benchmarks systematically evaluate VLMs' spatial reasoning capabilities in driving scenarios. To fill this gap, we propose NuScenes-SpatialQA, the first large-scale ground-truth-based Question-Answer (QA) benchmark specifically designed to evaluate the spatial understanding and reasoning capabilities of VLMs in autonomous driving. Built upon the NuScenes dataset, the benchmark is constructed through an automated 3D scene graph generation pipeline and a QA generation pipeline. The benchmark systematically evaluates VLMs' performance in both spatial understanding and reasoning across multiple dimensions. Using this benchmark, we conduct extensive experiments on diverse VLMs, including both general and spatial-enhanced models, providing the first comprehensive evaluation of their spatial capabilities in autonomous driving. Surprisingly, the experimental results show that the spatial-enhanced VLM outperforms in qualitative QA but does not demonstrate competitiveness in quantitative QA. In general, VLMs still face considerable challenges in spatial understanding and reasoning.
Abstract:Traditional diffusion models typically employ a U-Net architecture. Previous studies have unveiled the roles of attention blocks in the U-Net. However, they overlook the dynamic evolution of their importance during the inference process, which hinders their further exploitation to improve image applications. In this study, we first theoretically proved that, re-weighting the outputs of the Transformer blocks within the U-Net is a "free lunch" for improving the signal-to-noise ratio during the sampling process. Next, we proposed Importance Probe to uncover and quantify the dynamic shifts in importance of the Transformer blocks throughout the denoising process. Finally, we design an adaptive importance-based re-weighting schedule tailored to specific image generation and editing tasks. Experimental results demonstrate that, our approach significantly improves the efficiency of the inference process, and enhances the aesthetic quality of the samples with identity consistency. Our method can be seamlessly integrated into any U-Net-based architecture. Code: https://github.com/Hytidel/UNetReweighting
Abstract:This article addresses time-optimal path planning for a vehicle capable of moving both forward and backward on a unit sphere with a unit maximum speed, and constrained by a maximum absolute turning rate $U_{max}$. The proposed formulation can be utilized for optimal attitude control of underactuated satellites, optimal motion planning for spherical rolling robots, and optimal path planning for mobile robots on spherical surfaces or uneven terrains. By utilizing Pontryagin's Maximum Principle and analyzing phase portraits, it is shown that for $U_{max}\geq1$, the optimal path connecting a given initial configuration to a desired terminal configuration falls within a sufficient list of 23 path types, each comprising at most 6 segments. These segments belong to the set $\{C,G,T\}$, where $C$ represents a tight turn with radius $r=\frac{1}{\sqrt{1+U_{max}^2}}$, $G$ represents a great circular arc, and $T$ represents a turn-in-place motion. Closed-form expressions for the angles of each path in the sufficient list are derived. The source code for solving the time-optimal path problem and visualization is publicly available at https://github.com/sixuli97/Optimal-Spherical-Convexified-Reeds-Shepp-Paths.
Abstract:In multi-class unsupervised anomaly detection(MUAD), reconstruction-based methods learn to map input images to normal patterns to identify anomalous pixels. However, this strategy easily falls into the well-known "learning shortcut" issue when decoders fail to capture normal patterns and reconstruct both normal and abnormal samples naively. To address that, we propose to learn the input features in global and local manners, forcing the network to memorize the normal patterns more comprehensively. Specifically, we design a two-branch decoder block, named Omni-block. One branch corresponds to global feature learning, where we serialize two self-attention blocks but replace the query and (key, value) with learnable tokens, respectively, thus capturing global features of normal patterns concisely and thoroughly. The local branch comprises depth-separable convolutions, whose locality enables effective and efficient learning of local features for normal patterns. By stacking Omni-blocks, we build a framework, Omni-AD, to learn normal patterns of different granularity and reconstruct them progressively. Comprehensive experiments on public anomaly detection benchmarks show that our method outperforms state-of-the-art approaches in MUAD. Code is available at https://github.com/easyoo/Omni-AD.git.
Abstract:We present Video Motion Graphs, a system designed to generate realistic human motion videos. Using a reference video and conditional signals such as music or motion tags, the system synthesizes new videos by first retrieving video clips with gestures matching the conditions and then generating interpolation frames to seamlessly connect clip boundaries. The core of our approach is HMInterp, a robust Video Frame Interpolation (VFI) model that enables seamless interpolation of discontinuous frames, even for complex motion scenarios like dancing. HMInterp i) employs a dual-branch interpolation approach, combining a Motion Diffusion Model for human skeleton motion interpolation with a diffusion-based video frame interpolation model for final frame generation. ii) adopts condition progressive training to effectively leverage identity strong and weak conditions, such as images and pose. These designs ensure both high video texture quality and accurate motion trajectory. Results show that our Video Motion Graphs outperforms existing generative- and retrieval-based methods for multi-modal conditioned human motion video generation. Project page can be found at https://h-liu1997.github.io/Video-Motion-Graphs/
Abstract:The integration of geometric reconstruction and generative modeling remains a critical challenge in developing AI systems capable of human-like spatial reasoning. This paper proposes Aether, a unified framework that enables geometry-aware reasoning in world models by jointly optimizing three core capabilities: (1) 4D dynamic reconstruction, (2) action-conditioned video prediction, and (3) goal-conditioned visual planning. Through task-interleaved feature learning, Aether achieves synergistic knowledge sharing across reconstruction, prediction, and planning objectives. Building upon video generation models, our framework demonstrates unprecedented synthetic-to-real generalization despite never observing real-world data during training. Furthermore, our approach achieves zero-shot generalization in both action following and reconstruction tasks, thanks to its intrinsic geometric modeling. Remarkably, even without real-world data, its reconstruction performance is comparable with or even better than that of domain-specific models. Additionally, Aether employs camera trajectories as geometry-informed action spaces, enabling effective action-conditioned prediction and visual planning. We hope our work inspires the community to explore new frontiers in physically-reasonable world modeling and its applications.
Abstract:We introduce Visual Persona, a foundation model for text-to-image full-body human customization that, given a single in-the-wild human image, generates diverse images of the individual guided by text descriptions. Unlike prior methods that focus solely on preserving facial identity, our approach captures detailed full-body appearance, aligning with text descriptions for body structure and scene variations. Training this model requires large-scale paired human data, consisting of multiple images per individual with consistent full-body identities, which is notoriously difficult to obtain. To address this, we propose a data curation pipeline leveraging vision-language models to evaluate full-body appearance consistency, resulting in Visual Persona-500K, a dataset of 580k paired human images across 100k unique identities. For precise appearance transfer, we introduce a transformer encoder-decoder architecture adapted to a pre-trained text-to-image diffusion model, which augments the input image into distinct body regions, encodes these regions as local appearance features, and projects them into dense identity embeddings independently to condition the diffusion model for synthesizing customized images. Visual Persona consistently surpasses existing approaches, generating high-quality, customized images from in-the-wild inputs. Extensive ablation studies validate design choices, and we demonstrate the versatility of Visual Persona across various downstream tasks.
Abstract:Recent video diffusion models have enhanced video editing, but it remains challenging to handle instructional editing and diverse tasks (e.g., adding, removing, changing) within a unified framework. In this paper, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions, a simple end-to-end framework that unifies video concept editing, grounding, and reasoning based on diverse user instructions. Specifically, given a video and text query, VEGGIE first utilizes an MLLM to interpret user intentions in instructions and ground them to the video contexts, generating frame-specific grounded task queries for pixel-space responses. A diffusion model then renders these plans and generates edited videos that align with user intent. To support diverse tasks and complex instructions, we employ a curriculum learning strategy: first aligning the MLLM and video diffusion model with large-scale instructional image editing data, followed by end-to-end fine-tuning on high-quality multitask video data. Additionally, we introduce a novel data synthesis pipeline to generate paired instructional video editing data for model training. It transforms static image data into diverse, high-quality video editing samples by leveraging Image-to-Video models to inject dynamics. VEGGIE shows strong performance in instructional video editing with different editing skills, outperforming the best instructional baseline as a versatile model, while other models struggle with multi-tasking. VEGGIE also excels in video object grounding and reasoning segmentation, where other baselines fail. We further reveal how the multiple tasks help each other and highlight promising applications like zero-shot multimodal instructional and in-context video editing.
Abstract:Large foundation models trained on large-scale visual-text data can significantly enhance Open Vocabulary Object Detection (OVD) through data generation. However, this may lead to biased synthetic data and overfitting to specific configurations. It can sidestep biases of manually curated data generation by directly leveraging hidden states of Large Language Models (LLMs), which is surprisingly rarely explored. This paper presents a systematic method to enhance visual grounding by utilizing decoder layers of the LLM of a MLLM. We introduce a zero-initialized cross-attention adapter to enable efficient knowledge transfer from LLMs to object detectors, an new approach called LED (LLM Enhanced Open-Vocabulary Object Detection). We demonstrate that intermediate hidden states from early LLM layers retain strong spatial-semantic correlations that are beneficial to grounding tasks. Experiments show that our adaptation strategy significantly enhances the performance on complex free-form text queries while remaining the same on plain categories. With our adaptation, Qwen2-0.5B with Swin-T as the vision encoder improves GroundingDINO by 2.33% on Omnilabel, at the overhead of 8.7% more GFLOPs. Qwen2-0.5B with a larger vision encoder can further boost the performance by 6.22%. We further validate our design by ablating on varied adapter architectures, sizes of LLMs, and which layers to add adaptation.
Abstract:Recently, 2D Gaussian Splatting (2DGS) has demonstrated superior geometry reconstruction quality than the popular 3DGS by using 2D surfels to approximate thin surfaces. However, it falls short when dealing with glossy surfaces, resulting in visible holes in these areas. We found the reflection discontinuity causes the issue. To fit the jump from diffuse to specular reflection at different viewing angles, depth bias is introduced in the optimized Gaussian primitives. To address that, we first replace the depth distortion loss in 2DGS with a novel depth convergence loss, which imposes a strong constraint on depth continuity. Then, we rectified the depth criterion in determining the actual surface, which fully accounts for all the intersecting Gaussians along the ray. Qualitative and quantitative evaluations across various datasets reveal that our method significantly improves reconstruction quality, with more complete and accurate surfaces than 2DGS.