arXiv subjects

Zezhou Cheng

Publications and source records attributed to Zezhou Cheng.

At least 19 recordsLinked to original sources

arXiv2026

Foresight: Failure Detection for Long-Horizon Robotic Manipulation with Action-Conditioned World Model Latents

Long-horizon tasks are common in real-world robotic deployments, yet failure detection for such tasks remains underexplored. Detecting failures in long-horizon robotic tasks is particularly challenging because failure onset is often ambiguous and dense temporal annotations are typically unavailable. We present Foresight, a failure detection framework that monitors manipulation trajectories using latent representations from an action-conditioned world model. Foresight is trained using only final task-level success or failure labels. By leveraging predictive world-model embeddings, our method provides a unified framework for failure detection across different policies. We further use functional conformal prediction (FCP) to calibrate detection thresholds adaptively. We evaluate Foresight with state-of-the-art vision-language-action policies in simulation on LIBERO-Long, ManiSkill-Long, and BEHAVIOR-1K, compare it against state-of-the-artfailure detection methods, and validate it on real robots with three long-horizon tasks on a ReactorX-200 arm and one task on a Franka arm. Our results suggest that action-conditioned world-model embeddings provide a scalable representation for reliable failure monitoring in long-horizon manipulation.

cs.RO↗

arXiv2026

VLGA: Vision-Language-Geometry-Action Models for Autonomous Driving

Vision-language-action (VLA) models can describe scenes and reason about them in language, yet still struggle to ground their actions in the dense 3D world around them. Existing approaches either inject features from a frozen 3D foundation model without an objective that ensures the policy uses them, or constrain geometry with sparse box and map losses that provide no dense spatial signal. We introduce VLGA, the first vision-language-action model supervised to reconstruct the dense 3D world it drives through. VLGA introduces geometry as a fourth modality alongside vision, language, and action through a dedicated expert supervised by a per-pixel pointmap regression loss against LiDAR. Extensive experiments conducted on challenging nuScenes and Bench2Drive datasets for open-loop and closed-loop evaluations, respectively, show the superiority of VLGA over counterpart VLA methods. In particular, on open-loop nuScenes, VLGA sets a new state of the art among VLA methods without ego status, with the lowest L2 (0.50\,m average) and 3-second collision rate (0.18\%). On closed-loop Bench2Drive, VLGA attains the state-of-the-art driving score of 79.08, +0.71 over the strongest prior VLA, at comparable efficiency and comfort.

cs.CV↗

arXiv2026

OmniShotCut: Holistic Relational Shot Boundary Detection with Shot-Query Transformer

Shot Boundary Detection (SBD) aims to automatically identify shot changes and divide a video into coherent shots. While SBD was widely studied in the literature, existing methods often produce non-interpretable boundaries on transitions, miss subtle yet harmful discontinuities, and rely on noisy, low-diversity annotations and outdated benchmarks. To alleviate these limitations, we propose OmniShotCut to formulate SBD as structured relational prediction, jointly estimating shot ranges with intra-shot relations and inter-shot relations, by a shot query-based dense video Transformer. To avoid imprecise manual labeling, we adopt a fully synthetic transition synthesis pipeline that automatically reproduces major transition families with precise boundaries and parameterized variants. We also introduce OmniShotCutBench, a modern wide-domain benchmark enabling holistic and diagnostic evaluation. Experiments on the benchmarks demonstrate the effectiveness and generality of our method.

cs.CV↗

arXiv2026

WildRayZer: Self-supervised Large View Synthesis in Dynamic Environments

We present WildRayZer, a self-supervised framework for novel view synthesis (NVS) in dynamic environments where both the camera and objects move. Dynamic content breaks the multi-view consistency that static NVS models rely on, leading to ghosting, hallucinated geometry, and unstable pose estimation. WildRayZer addresses this by performing an analysis-by-synthesis test: a camera-only static renderer explains rigid structure, and its residuals reveal transient regions. From these residuals, we construct pseudo motion masks, distill a motion estimator, and use it to mask input tokens and gate loss gradients so supervision focuses on cross-view background completion. To enable large-scale training and evaluation, we curate Dynamic RealEstate10K (D-RE10K), a real-world dataset of 15K casually captured dynamic sequences, and D-RE10K-iPhone, a paired transient and clean benchmark for sparse-view transient-aware NVS. Experiments show that WildRayZer consistently outperforms optimization-based and feed-forward baselines in both transient-region removal and full-frame NVS quality with a single feed-forward pass.

cs.CV↗

arXiv2026

LabelAny3D: Label Any Object 3D in the Wild

Detecting objects in 3D space from monocular input is crucial for applications ranging from robotics to scene understanding. Despite advanced performance in the indoor and autonomous driving domains, existing monocular 3D detection models struggle with in-the-wild images due to the lack of 3D in-the-wild datasets and the challenges of 3D annotation. We introduce LabelAny3D, an \emph{analysis-by-synthesis} framework that reconstructs holistic 3D scenes from 2D images to efficiently produce high-quality 3D bounding box annotations. Built on this pipeline, we present COCO3D, a new benchmark for open-vocabulary monocular 3D detection, derived from the MS-COCO dataset and covering a wide range of object categories absent from existing 3D datasets. Experiments show that annotations generated by LabelAny3D improve monocular 3D detection performance across multiple benchmarks, outperforming prior auto-labeling approaches in quality. These results demonstrate the promise of foundation-model-driven annotation for scaling up 3D recognition in realistic, open-world settings.

cs.CV↗

arXiv2025

Empowering Dynamic Urban Navigation with Stereo and Mid-Level Vision

The success of foundation models in language and vision motivated research in fully end-to-end robot navigation foundation models (NFMs). NFMs directly map monocular visual input to control actions and ignore mid-level vision modules (tracking, depth estimation, etc) entirely. While the assumption that vision capabilities will emerge implicitly is compelling, it requires large amounts of pixel-to-action supervision that are difficult to obtain. The challenge is especially pronounced in dynamic and unstructured settings, where robust navigation requires precise geometric and dynamic understanding, while the depth-scale ambiguity in monocular views further limits accurate spatial reasoning. In this paper, we show that relying on monocular vision and ignoring mid-level vision priors is inefficient. We present StereoWalker, which augments NFMs with stereo inputs and explicit mid-level vision such as depth estimation and dense pixel tracking. Our intuition is straightforward: stereo inputs resolve the depth-scale ambiguity, and modern mid-level vision models provide reliable geometric and motion structure in dynamic scenes. We also curate a large stereo navigation dataset with automatic action annotation from Internet stereo videos to support training of StereoWalker and to facilitate future research. Through our experiments, we find that mid-level vision enables StereoWalker to achieve a comparable performance as the state-of-the-art using only 1.5% of the training data, and surpasses the state-of-the-art using the full data. We also observe that stereo vision yields higher navigation performance than monocular input.

cs.CV↗

arXiv2025

SAB3R: Semantic-Augmented Backbone in 3D Reconstruction

We introduce a new task, Map and Locate, which unifies the traditionally distinct objectives of open-vocabulary segmentation - detecting and segmenting object instances based on natural language queries - and 3D reconstruction, the process of estimating a scene's 3D structure from visual inputs. Specifically, Map and Locate involves generating a point cloud from an unposed video and segmenting object instances based on open-vocabulary queries. This task serves as a critical step toward real-world embodied AI applications and introduces a practical task that bridges reconstruction, recognition and reorganization. To tackle this task, we introduce a simple yet effective baseline, which we denote as SAB3R. Our approach builds upon MASt3R, a recent breakthrough in 3D computer vision, and incorporates a lightweight distillation strategy. This method transfers dense, per-pixel semantic features from 2D vision backbones (eg, CLIP and DINOv2) to enhance MASt3R's capabilities. Without introducing any auxiliary frozen networks, our model generates per-pixel semantic features and constructs cohesive point maps in a single forward pass. Compared to separately deploying MASt3R and CLIP, our unified model, SAB3R, achieves superior performance on the Map and Locate benchmark. Furthermore, we evaluate SAB3R on both 2D semantic segmentation and 3D tasks to comprehensively validate its effectiveness.

cs.CV↗

arXiv2025

Point-MoE: Large-Scale Multi-Dataset Training with Mixture-of-Experts for 3D Semantic Segmentation

While massively scaling both data and models have become central in NLP and 2D vision, their benefits for 3D point cloud understanding remain limited. We study the initial step of scaling 3D point cloud understanding under a realistic regime: large-scale multi-dataset joint training for 3D semantic segmentation, with no dataset labels available at training or inference time. Point clouds arise from a wide range of sensors (e.g., depth cameras, LiDAR) and scenes (\eg, indoor, outdoor), yielding heterogeneous scanning patterns, sampling densities, and semantic biases; naively mixing such datasets degrades standard models. Therefore, we introduce Point-MoE, a Mixture-of-Experts design that expands model capacity through sparsely activated expert MLPs and a lightweight top-$k$ router, allowing tokens to select specialized experts without requiring dataset supervision. Trained jointly on a diverse mix of indoor and outdoor datasets, and evaluated on seen datasets as well as in zero-shot settings, Point-MoE outperforms prior methods without using dataset labels for either training or inference. This outlines a scalable path for 3D perception: letting the model discover structure in heterogeneous 3D data rather than imposing it via manual curation or dataset-specific heuristics.

cs.CV↗

arXiv2025

Frame In-N-Out: Unbounded Controllable Image-to-Video Generation

Controllability, temporal coherence, and detail synthesis remain the most critical challenges in video generation. In this paper, we focus on a commonly used yet underexplored cinematic technique known as Frame In and Frame Out. Specifically, starting from image-to-video generation, users can control the objects in the image to naturally leave the scene or provide breaking new identity references to enter the scene, guided by a user-specified motion trajectory. To support this task, we introduce a new dataset that is curated semi-automatically, an efficient identity-preserving motion-controllable video Diffusion Transformer architecture, and a comprehensive evaluation protocol targeting this task. Our evaluation shows that our proposed approach significantly outperforms existing baselines.

cs.CV↗

arXiv2024

Open Vocabulary Monocular 3D Object Detection

We propose and study open-vocabulary monocular 3D detection, a novel task that aims to detect objects of any categores in metric 3D space from a single RGB image. Existing 3D object detectors either rely on costly sensors such as LiDAR or multi-view setups, or remain confined to closed vocabularies settings with limited categories, restricting their applicability. We identify two key challenges in this new setting. First, the scarcity of 3D bounding box annotations limits the ability to train generalizable models. To reduce dependence on 3D supervision, we propose a framework that effectively integrates pretrained 2D and 3D vision foundation models. Second, missing labels and semantic ambiguities (\eg, table vs. desk) in existing datasets hinder reliable evaluation. To address this, we design a novel metric that captures model performance while mitigating annotation issues. Our approach achieves state-of-the-art results in zero-shot 3D detection of novel categories as well as in-domain detection on seen classes. We hope our method provides a strong baseline and our evaluation protocol establishes a reliable benchmark for future research.

cs.CV↗

arXiv2024

Semantic-Free Procedural 3D Shapes Are Surprisingly Good Teachers

Self-supervised learning has emerged as a promising approach for acquiring transferable 3D representations from unlabeled 3D point clouds. Unlike 2D images, which are widely accessible, acquiring 3D assets requires specialized expertise or professional 3D scanning equipment, making it difficult to scale and raising copyright concerns. To address these challenges, we propose learning 3D representations from procedural 3D programs that automatically generate 3D shapes using simple 3D primitives and augmentations. Remarkably, despite lacking semantic content, the 3D representations learned from the procedurally generated 3D shapes perform on par with state-of-the-art representations learned from semantically recognizable 3D models (e.g., airplanes) across various downstream 3D tasks, such as shape classification, part segmentation, masked point cloud completion, and both scene semantic and instance segmentation. We provide a detailed analysis on factors that make a good 3D procedural programs. Extensive experiments further suggest that current 3D self-supervised learning methods on point clouds do not rely on semantics of 3D shapes, shedding light on the nature of 3D representations learned.

cs.CV↗

arXiv2024

Probing the Mid-level Vision Capabilities of Self-Supervised Learning

Mid-level vision capabilities - such as generic object localization and 3D geometric understanding - are not only fundamental to human vision but are also crucial for many real-world applications of computer vision. These abilities emerge with minimal supervision during the early stages of human visual development. Despite their significance, current self-supervised learning (SSL) approaches are primarily designed and evaluated for high-level recognition tasks, leaving their mid-level vision capabilities largely unexamined. In this study, we introduce a suite of benchmark protocols to systematically assess mid-level vision capabilities and present a comprehensive, controlled evaluation of 22 prominent SSL models across 8 mid-level vision tasks. Our experiments reveal a weak correlation between mid-level and high-level task performance. We also identify several SSL methods with highly imbalanced performance across mid-level and high-level capabilities, as well as some that excel in both. Additionally, we investigate key factors contributing to mid-level vision performance, such as pretraining objectives and network architectures. Our study provides a holistic and timely view of what SSL models have learned, complementing existing research that primarily focuses on high-level vision tasks. We hope our findings guide future SSL research to benchmark models not only on high-level vision tasks but on mid-level as well.

cs.CV↗

arXiv2024

Machine Unlearning of Pre-trained Large Language Models

This study investigates the concept of the `right to be forgotten' within the context of large language models (LLMs). We explore machine unlearning as a pivotal solution, with a focus on pre-trained models--a notably under-researched area. Our research delineates a comprehensive framework for machine unlearning in pre-trained LLMs, encompassing a critical analysis of seven diverse unlearning methods. Through rigorous evaluation using curated datasets from arXiv, books, and GitHub, we establish a robust benchmark for unlearning performance, demonstrating that these methods are over $10^5$ times more computationally efficient than retraining. Our results show that integrating gradient ascent with gradient descent on in-distribution data improves hyperparameter robustness. We also provide detailed guidelines for efficient hyperparameter tuning in the unlearning process. Our findings advance the discourse on ethical AI practices, offering substantive insights into the mechanics of machine unlearning for pre-trained LLMs and underscoring the potential for responsible AI development.

cs.CL↗

arXiv2023

LU-NeRF: Scene and Pose Estimation by Synchronizing Local Unposed NeRFs

A critical obstacle preventing NeRF models from being deployed broadly in the wild is their reliance on accurate camera poses. Consequently, there is growing interest in extending NeRF models to jointly optimize camera poses and scene representation, which offers an alternative to off-the-shelf SfM pipelines which have well-understood failure modes. Existing approaches for unposed NeRF operate under limited assumptions, such as a prior pose distribution or coarse pose initialization, making them less effective in a general setting. In this work, we propose a novel approach, LU-NeRF, that jointly estimates camera poses and neural radiance fields with relaxed assumptions on pose configuration. Our approach operates in a local-to-global manner, where we first optimize over local subsets of the data, dubbed mini-scenes. LU-NeRF estimates local pose and geometry for this challenging few-shot task. The mini-scene poses are brought into a global reference frame through a robust pose synchronization step, where a final global optimization of pose and scene can be performed. We show our LU-NeRF pipeline outperforms prior attempts at unposed NeRF without making restrictive assumptions on the pose prior. This allows us to operate in the general SE(3) pose setting, unlike the baselines. Our results also indicate our model can be complementary to feature-based SfM pipelines as it compares favorably to COLMAP on low-texture and low-resolution images.

cs.CV↗

arXiv2022

Accidental Turntables: Learning 3D Pose by Watching Objects Turn

We propose a technique for learning single-view 3D object pose estimation models by utilizing a new source of data -- in-the-wild videos where objects turn. Such videos are prevalent in practice (e.g., cars in roundabouts, airplanes near runways) and easy to collect. We show that classical structure-from-motion algorithms, coupled with the recent advances in instance detection and feature matching, provides surprisingly accurate relative 3D pose estimation on such videos. We propose a multi-stage training scheme that first learns a canonical pose across a collection of videos and then supervises a model for single-view pose estimation. The proposed technique achieves competitive performance with respect to existing state-of-the-art on standard benchmarks for 3D pose estimation, without requiring any pose labels during training. We also contribute an Accidental Turntables Dataset, containing a challenging set of 41,212 images of cars in cluttered backgrounds, motion blur and illumination changes that serves as a benchmark for 3D pose estimation.

cs.CV↗

arXiv2022

Improving Few-Shot Part Segmentation using Coarse Supervision

A significant bottleneck in training deep networks for part segmentation is the cost of obtaining detailed annotations. We propose a framework to exploit coarse labels such as figure-ground masks and keypoint locations that are readily available for some categories to improve part segmentation models. A key challenge is that these annotations were collected for different tasks and with different labeling styles and cannot be readily mapped to the part labels. To this end, we propose to jointly learn the dependencies between labeling styles and the part segmentation model, allowing us to utilize supervision from diverse labels. To evaluate our approach we develop a benchmark on the Caltech-UCSD birds and OID Aircraft dataset. Our approach outperforms baselines based on multi-task learning, semi-supervised learning, and competitive methods relying on loss functions manually designed to exploit sparse-supervision.

cs.CV↗

arXiv2022

Cross-Modal 3D Shape Generation and Manipulation

Creating and editing the shape and color of 3D objects require tremendous human effort and expertise. Compared to direct manipulation in 3D interfaces, 2D interactions such as sketches and scribbles are usually much more natural and intuitive for the users. In this paper, we propose a generic multi-modal generative model that couples the 2D modalities and implicit 3D representations through shared latent spaces. With the proposed model, versatile 3D generation and manipulation are enabled by simply propagating the editing from a specific 2D controlling modality through the latent spaces. For example, editing the 3D shape by drawing a sketch, re-colorizing the 3D surface via painting color scribbles on the 2D rendering, or generating 3D shapes of a certain category given one or a few reference images. Unlike prior works, our model does not require re-training or fine-tuning per editing task and is also conceptually simple, easy to implement, robust to input domain shifts, and flexible to diverse reconstruction on partial 2D inputs. We evaluate our framework on two representative 2D modalities of grayscale line sketches and rendered color images, and demonstrate that our method enables various shape manipulation and generation tasks with these 2D modalities.

cs.CV↗

arXiv2022

GANORCON: Are Generative Models Useful for Few-shot Segmentation?

Advances in generative modeling based on GANs has motivated the community to find their use beyond image generation and editing tasks. In particular, several recent works have shown that GAN representations can be re-purposed for discriminative tasks such as part segmentation, especially when training data is limited. But how do these improvements stack-up against recent advances in self-supervised learning? Motivated by this we present an alternative approach based on contrastive learning and compare their performance on standard few-shot part segmentation benchmarks. Our experiments reveal that not only do the GAN-based approach offer no significant performance advantage, their multi-step training is complex, nearly an order-of-magnitude slower, and can introduce additional bias. These experiments suggest that the inductive biases of generative models, such as their ability to disentangle shape and texture, are well captured by standard feed-forward networks trained using contrastive learning. These experiments suggest that the inductive biases present in current generative models, such as their ability to disentangle shape and texture, are well captured by standard feed-forward networks trained using contrastive learning.

cs.CV↗