Ian Reid

Libo Sun, Jia-Wang Bian, Huangying Zhan et al. · bytedance, oxford

Self-supervised monocular depth estimation has shown impressive results in static scenes. It relies on the multi-view consistency assumption for training networks, however, that is violated in dynamic object regions and occlusions. Consequently, existing methods show poor accuracy in dynamic scenes, and the estimated depth map is blurred at object boundaries because they are usually occluded in other training views. In this paper, we propose SC-DepthV3 for addressing the challenges. Specifically, we introduce an external pretrained monocular depth estimation model for generating single-image depth prior, namely pseudo-depth, based on which we propose novel losses to boost self-supervised training. As a result, our model can predict sharp and accurate depth maps, even when training from monocular videos of highly-dynamic scenes. We demonstrate the significantly superior performance of our method over previous methods on six challenging datasets, and we provide detailed ablation studies for the proposed terms. Source code and data will be released at https://github.com/JiawangBian/sc_depth_pl

Junlin Han, Lars Petersson, Hongdong Li et al. · oxford

We present a simple method, CropMix, for the purpose of producing a rich input distribution from the original dataset distribution. Unlike single random cropping, which may inadvertently capture only limited information, or irrelevant information, like pure background, unrelated objects, etc, we crop an image multiple times using distinct crop scales, thereby ensuring that multi-scale information is captured. The new input distribution, serving as training data, useful for a number of vision tasks, is then formed by simply mixing multiple cropped views. We first demonstrate that CropMix can be seamlessly applied to virtually any training recipe and neural network architecture performing classification tasks. CropMix is shown to improve the performance of image classifiers on several benchmark tasks across-the-board without sacrificing computational simplicity and efficiency. Moreover, we show that CropMix is of benefit to both contrastive learning and masked image modeling towards more powerful representations, where preferable results are achieved when learned representations are transferred to downstream tasks. Code is available at GitHub.

Yuyuan Liu, Choubo Ding, Yu Tian et al.

Semantic segmentation models classify pixels into a set of known (``in-distribution'') visual classes. When deployed in an open world, the reliability of these models depends on their ability not only to classify in-distribution pixels but also to detect out-of-distribution (OoD) pixels. Historically, the poor OoD detection performance of these models has motivated the design of methods based on model re-training using synthetic training images that include OoD visual objects. Although successful, these re-trained methods have two issues: 1) their in-distribution segmentation accuracy may drop during re-training, and 2) their OoD detection accuracy does not generalise well to new contexts (e.g., country surroundings) outside the training set (e.g., city surroundings). In this paper, we mitigate these issues with: (i) a new residual pattern learning (RPL) module that assists the segmentation model to detect OoD pixels without affecting the inlier segmentation performance; and (ii) a novel context-robust contrastive learning (CoroCL) that enforces RPL to robustly detect OoD pixels among various contexts. Our approach improves by around 10\% FPR and 7\% AuPRC the previous state-of-the-art in Fishyscapes, Segment-Me-If-You-Can, and RoadAnomaly datasets. Our code is available at: https://github.com/yyliu01/RPL.

Anh-Dzung Doan, Bach Long Nguyen, Surabhi Gupta et al.

To ensure reliable object detection in autonomous systems, the detector must be able to adapt to changes in appearance caused by environmental factors such as time of day, weather, and seasons. Continually adapting the detector to incorporate these changes is a promising solution, but it can be computationally costly. Our proposed approach is to selectively adapt the detector only when necessary, using new data that does not have the same distribution as the current training data. To this end, we investigate three popular metrics for domain gap evaluation and find that there is a correlation between the domain gap and detection accuracy. Therefore, we apply the domain gap as a criterion to decide when to adapt the detector. Our experiments show that our approach has the potential to improve the efficiency of the detector's operation in real-world scenarios, where environmental conditions change in a cyclical manner, without sacrificing the overall performance of the detector. Our code is publicly available at https://github.com/dadung/DGE-CDA.

Krishan Rana, Jesse Haviland, Sourav Garg et al.

Large language models (LLMs) have demonstrated impressive results in developing generalist planning agents for diverse tasks. However, grounding these plans in expansive, multi-floor, and multi-room environments presents a significant challenge for robotics. We introduce SayPlan, a scalable approach to LLM-based, large-scale task planning for robotics using 3D scene graph (3DSG) representations. To ensure the scalability of our approach, we: (1) exploit the hierarchical nature of 3DSGs to allow LLMs to conduct a 'semantic search' for task-relevant subgraphs from a smaller, collapsed representation of the full graph; (2) reduce the planning horizon for the LLM by integrating a classical path planner and (3) introduce an 'iterative replanning' pipeline that refines the initial plan using feedback from a scene graph simulator, correcting infeasible actions and avoiding planning failures. We evaluate our approach on two large-scale environments spanning up to 3 floors and 36 rooms with 140 assets and objects and show that our approach is capable of grounding large-scale, long-horizon task plans from abstract, and natural language instruction for a mobile manipulator robot to execute. We provide real robot video demonstrations on our project page https://sayplan.github.io.

Huangying Zhan, Jiyang Zheng, Yi Xu et al.

A high-quality 3D reconstruction of a scene from a collection of 2D images can be achieved through offline/online mapping methods. In this paper, we explore active mapping from the perspective of implicit representations, which have recently produced compelling results in a variety of applications. One of the most popular implicit representations - Neural Radiance Field (NeRF), first demonstrated photorealistic rendering results using multi-layer perceptrons, with promising offline 3D reconstruction as a by-product of the radiance field. More recently, researchers also applied this implicit representation for online reconstruction and localization (i.e. implicit SLAM systems). However, the study on using implicit representation for active vision tasks is still very limited. In this paper, we are particularly interested in applying the neural radiance field for active mapping and planning problems, which are closely coupled tasks in an active system. We, for the first time, present an RGB-only active vision framework using radiance field representation for active 3D reconstruction and planning in an online manner. Specifically, we formulate this joint task as an iterative dual-stage optimization problem, where we alternatively optimize for the radiance field representation and path planning. Experimental results suggest that the proposed method achieves competitive results compared to other offline methods and outperforms active reconstruction methods using NeRFs.

Boying Li, Zhixi Cai, Yuan-Fang Li et al.

We propose Hier-SLAM, a semantic 3D Gaussian Splatting SLAM method featuring a novel hierarchical categorical representation, which enables accurate global 3D semantic mapping, scaling-up capability, and explicit semantic label prediction in the 3D world. The parameter usage in semantic SLAM systems increases significantly with the growing complexity of the environment, making it particularly challenging and costly for scene understanding. To address this problem, we introduce a novel hierarchical representation that encodes semantic information in a compact form into 3D Gaussian Splatting, leveraging the capabilities of large language models (LLMs). We further introduce a novel semantic loss designed to optimize hierarchical semantic information through both inter-level and cross-level optimization. Furthermore, we enhance the whole SLAM system, resulting in improved tracking and mapping performance. Our \MethodName{} outperforms existing dense SLAM methods in both mapping and tracking accuracy, while achieving a 2x operation speed-up. Additionally, it achieves on-par semantic rendering performance compared to existing methods while significantly reducing storage and training time requirements. Rendering FPS impressively reaches 2,000 with semantic information and 3,000 without it. Most notably, it showcases the capability of handling the complex real-world scene with more than 500 semantic classes, highlighting its valuable scaling-up capability. The open-source code is available at https://github.com/LeeBY68/Hier-SLAM

Junlin Han, Huangying Zhan, Jie Hong et al. · oxford

This paper studies the problem of measuring and predicting how memorable an image is to pattern recognition machines, as a path to explore machine intelligence. Firstly, we propose a self-supervised machine memory quantification pipeline, dubbed ``MachineMem measurer'', to collect machine memorability scores of images. Similar to humans, machines also tend to memorize certain kinds of images, whereas the types of images that machines and humans memorize are different. Through in-depth analysis and comprehensive visualizations, we gradually unveil that``complex" images are usually more memorable to machines. We further conduct extensive experiments across 11 different machines (from linear classifiers to modern ViTs) and 9 pre-training methods to analyze and understand machine memory. This work proposes the concept of machine memorability and opens a new research direction at the interface between machine memory and visual data.

Yuyuan Liu, Yuanhong Chen, Hu Wang et al.

The costly and time-consuming annotation process to produce large training sets for modelling semantic LiDAR segmentation methods has motivated the development of semi-supervised learning (SSL) methods. However, such SSL approaches often concentrate on employing consistency learning only for individual LiDAR representations. This narrow focus results in limited perturbations that generally fail to enable effective consistency learning. Additionally, these SSL approaches employ contrastive learning based on the sampling from a limited set of positive and negative embedding samples. This paper introduces a novel semi-supervised LiDAR semantic segmentation framework called ItTakesTwo (IT2). IT2 is designed to ensure consistent predictions from peer LiDAR representations, thereby improving the perturbation effectiveness in consistency learning. Furthermore, our contrastive learning employs informative samples drawn from a distribution of positive and negative embeddings learned from the entire training set. Results on public benchmarks show that our approach achieves remarkable improvements over the previous state-of-the-art (SOTA) methods in the field. The code is available at: https://github.com/yyliu01/IT2.

Ryan Faulkner, Luke Haub, Simon Ratcliffe et al.

LiDAR scanning for surveying applications acquire measurements over wide areas and long distances, which produces large-scale 3D point clouds with significant local density variations. While existing 3D semantic segmentation models conduct downsampling and upsampling to build robustness against varying point densities, they are less effective under the large local density variations characteristic of point clouds from surveying applications. To alleviate this weakness, we propose a novel architecture called HDVNet that contains a nested set of encoder-decoder pathways, each handling a specific point density range. Limiting the interconnections between the feature maps enables HDVNet to gauge the reliability of each feature based on the density of a point, e.g., downweighting high density features not existing in low density objects. By effectively handling input density variations, HDVNet outperforms state-of-the-art models in segmentation accuracy on real point clouds with inconsistent density, using just over half the weights.

Daqi Liu, Alvaro Parra, Yasir Latif et al.

Event cameras open up new possibilities for robotic perception due to their low latency and high dynamic range. On the other hand, developing effective event-based vision algorithms that fully exploit the beneficial properties of event cameras remains work in progress. In this paper, we focus on event-based visual odometry (VO). While existing event-driven VO pipelines have adopted continuous-time representations to asynchronously process event data, they either assume a known map, restrict the camera to planar trajectories, or integrate other sensors into the system. Towards map-free event-only monocular VO in SE(3), we propose an asynchronous structure-from-motion optimisation back-end. Our formulation is underpinned by a principled joint optimisation problem involving non-parametric Gaussian Process motion modelling and incremental maximum a posteriori inference. A high-performance incremental computation engine is employed to reason about the camera trajectory with every incoming event. We demonstrate the robustness of our asynchronous back-end in comparison to frame-based methods which depend on accurate temporal accumulation of measurements.

Sofia McLeod, Gabriele Meoni, Dario Izzo et al.

Event sensing is a major component in bio-inspired flight guidance and control systems. We explore the usage of event cameras for predicting time-to-contact (TTC) with the surface during ventral landing. This is achieved by estimating divergence (inverse TTC), which is the rate of radial optic flow, from the event stream generated during landing. Our core contributions are a novel contrast maximisation formulation for event-based divergence estimation, and a branch-and-bound algorithm to exactly maximise contrast and find the optimal divergence value. GPU acceleration is conducted to speed up the global algorithm. Another contribution is a new dataset containing real event streams from ventral landing that was employed to test and benchmark our method. Owing to global optimisation, our algorithm is much more capable at recovering the true divergence, compared to other heuristic divergence estimators or event-based optic flow methods. With GPU acceleration, our method also achieves competitive runtimes.

Huangying Zhan, Hamid Rezatofighi, Ian Reid

We propose a robotic learning system for autonomous exploration and navigation in unexplored environments. We are motivated by the idea that even an unseen environment may be familiar from previous experiences in similar environments. The core of our method, therefore, is a process for building, predicting, and using probabilistic layout graphs for assisting goal-based visual navigation. We describe a navigation system that uses the layout predictions to satisfy high-level goals (e.g. "go to the kitchen") more rapidly and accurately than the prior art. Our proposed navigation framework comprises three stages: (1) Perception and Mapping: building a multi-level 3D scene graph; (2) Prediction: predicting probabilistic 3D scene graph for the unexplored environment; (3) Navigation: assisting navigation with the graphs. We test our framework in Matterport3D and show more success and efficient navigation in unseen environments.

Raphael Lafargue, Luke Smith, Franck Vermet et al.

The predominant method for computing confidence intervals (CI) in few-shot learning (FSL) is based on sampling the tasks with replacement, i.e.\ allowing the same samples to appear in multiple tasks. This makes the CI misleading in that it takes into account the randomness of the sampler but not the data itself. To quantify the extent of this problem, we conduct a comparative analysis between CIs computed with and without replacement. These reveal a notable underestimation by the predominant method. This observation calls for a reevaluation of how we interpret confidence intervals and the resulting conclusions in FSL comparative studies. Our research demonstrates that the use of paired tests can partially address this issue. Additionally, we explore methods to further reduce the (size of the) CI by strategically sampling tasks of a specific size. We also introduce a new optimized benchmark, which can be accessed at https://github.com/RafLaf/FSL-benchmark-again

Anh-Dzung Doan, Vu Minh Hieu Phan, Surabhi Gupta et al.

Infrared imaging offers resilience against changing lighting conditions by capturing object temperatures. Yet, in few scenarios, its lack of visual details compared to daytime visible images, poses a significant challenge for human and machine interpretation. This paper proposes a novel diffusion method, dubbed Temporally Consistent Patch Diffusion Models (TC-DPM), for infrared-to-visible video translation. Our method, extending the Patch Diffusion Model, consists of two key components. Firstly, we propose a semantic-guided denoising, leveraging the strong representations of foundational models. As such, our method faithfully preserves the semantic structure of generated visible images. Secondly, we propose a novel temporal blending module to guide the denoising trajectory, ensuring the temporal consistency between consecutive frames. Experiment shows that TC-PDM outperforms state-of-the-art methods by 35.3% in FVD for infrared-to-visible video translation and by 6.1% in AP50 for day-to-night object detection. Our code is publicly available at https://github.com/dzungdoan6/tc-pdm

Anh-Dzung Doan, Bach Long Nguyen, Terry Lim et al.

Prior to deployment, an object detector is trained on a dataset compiled from a previous data collection campaign. However, the environment in which the object detector is deployed will invariably evolve, particularly in outdoor settings where changes in lighting, weather and seasons will significantly affect the appearance of the scene and target objects. It is almost impossible for all potential scenarios that the object detector may come across to be present in a finite training dataset. This necessitates continuous updates to the object detector to maintain satisfactory performance. Test-time domain adaptation techniques enable machine learning models to self-adapt based on the distributions of the testing data. However, existing methods mainly focus on fully automated adaptation, which makes sense for applications such as self-driving cars. Despite the prevalence of fully automated approaches, in some applications such as surveillance, there is usually a human operator overseeing the system's operation. We propose to involve the operator in test-time domain adaptation to raise the performance of object detection beyond what is achievable by fully automated adaptation. To reduce manual effort, the proposed method only requires the operator to provide weak labels, which are then used to guide the adaptation process. Furthermore, the proposed method can be performed in a streaming setting, where each online sample is observed only once. We show that the proposed method outperforms existing works, demonstrating a great benefit of human-in-the-loop test-time domain adaptation. Our code is publicly available at https://github.com/dzungdoan6/WSTTA

Bahey Tharwat, Giorgos Kordopatis-Zilos, Pavel Suma et al.

Multimodal Large Language Models (MLLMs) have demonstrated strong cross-modal reasoning capabilities, yet their potential for vision-only tasks remains underexplored. We investigate MLLMs as training-free similarity estimators for instance-level image-to-image retrieval. Our approach prompts the model with paired images and converts next-token probabilities into similarity scores, enabling zero-shot re-ranking within large-scale retrieval pipelines. This design avoids specialized architectures and fine-tuning, leveraging the rich visual discrimination learned during multimodal pre-training. We address scalability by combining MLLMs with memory-efficient indexing and top-$k$ candidate re-ranking. Experiments across diverse benchmarks show that MLLMs outperform task-specific re-rankers outside their native domains and exhibit superior robustness to clutter, occlusion, and small objects. Despite strong results, we identify failure modes under severe appearance changes, highlighting opportunities for future research. Our findings position MLLMs as a promising alternative for open-world large-scale image retrieval.

Abdelrahman Shaker, Ahmed Heakl, Jaseel Muhammad et al.

Unified multimodal models can both understand and generate visual content within a single architecture. Existing models, however, remain data-hungry and too heavy for deployment on edge devices. We present Mobile-O, a compact vision-language-diffusion model that brings unified multimodal intelligence to a mobile device. Its core module, the Mobile Conditioning Projector (MCP), fuses vision-language features with a diffusion generator using depthwise-separable convolutions and layerwise alignment. This design enables efficient cross-modal conditioning with minimal computational cost. Trained on only a few million samples and post-trained in a novel quadruplet format (generation prompt, image, question, answer), Mobile-O jointly enhances both visual understanding and generation capabilities. Despite its efficiency, Mobile-O attains competitive or superior performance compared to other unified models, achieving 74% on GenEval and outperforming Show-O and JanusFlow by 5% and 11%, while running 6x and 11x faster, respectively. For visual understanding, Mobile-O surpasses them by 15.3% and 5.1% averaged across seven benchmarks. Running in only ~3s per 512x512 image on an iPhone, Mobile-O establishes the first practical framework for real-time unified multimodal understanding and generation on edge devices. We hope Mobile-O will ease future research in real-time unified multimodal intelligence running entirely on-device with no cloud dependency. Our code, models, datasets, and mobile application are publicly available at https://amshaker.github.io/Mobile-O/

Zeyu Zhang, Akide Liu, Qi Chen et al.

Text-to-motion generation holds potential for film, gaming, and robotics, yet current methods often prioritize short motion generation, making it challenging to produce long motion sequences effectively: (1) Current methods struggle to handle long motion sequences as a single input due to prohibitively high computational cost; (2) Breaking down the generation of long motion sequences into shorter segments can result in inconsistent transitions and requires interpolation or inpainting, which lacks entire sequence modeling. To solve these challenges, we propose InfiniMotion, a method that generates continuous motion sequences of arbitrary length within an autoregressive framework. We highlight its groundbreaking capability by generating a continuous 1-hour human motion with around 80,000 frames. Specifically, we introduce the Motion Memory Transformer with Bidirectional Mamba Memory, enhancing the transformer's memory to process long motion sequences effectively without overwhelming computational resources. Notably our method achieves over 30% improvement in FID and 6 times longer demonstration compared to previous state-of-the-art methods, showcasing significant advancements in long motion generation. See project webpage: https://steve-zeyu-zhang.github.io/InfiniMotion/

Huy Hoang Nguyen, An Vuong, Anh Nguyen et al.

Grasp detection is a fundamental robotic task critical to the success of many industrial applications. However, current language-driven models for this task often struggle with cluttered images, lengthy textual descriptions, or slow inference speed. We introduce GraspMamba, a new language-driven grasp detection method that employs hierarchical feature fusion with Mamba vision to tackle these challenges. By leveraging rich visual features of the Mamba-based backbone alongside textual information, our approach effectively enhances the fusion of multimodal features. GraspMamba represents the first Mamba-based grasp detection model to extract vision and language features at multiple scales, delivering robust performance and rapid inference time. Intensive experiments show that GraspMamba outperforms recent methods by a clear margin. We validate our approach through real-world robotic experiments, highlighting its fast inference speed.

An Dinh Vuong, Tuan Van Vo, Abdullah Sohail et al.

World Models (WMs) have emerged as a promising approach for post-training Vision-Language-Action (VLA) policies to improve robustness and generalization under environmental changes. However, most WM-based post-training methods rely on pixel-space supervision, making policies sensitive to pixel-level artifacts and hallucination from imperfect WM rollouts. We introduce World2Act, a post-training framework that aligns VLA actions directly with WM video-dynamics latents using a contrastive matching objective, reducing dependence on pixels. Post-training performance is tied to rollout quality, yet current WMs struggle with arbitrary-length video generation as they are mostly trained on fixed-length clips while robotic execution durations vary widely. To address this, we propose an automatic LLM-based skill-decomposition pipeline that segments high-level instructions into low-level prompts. Our pipeline produces RoboCasa-Skill and LIBERO-Skill, supporting skill-compositional WMs that remain temporally consistent across diverse task horizons. Empirically, applying World2Act to VLAs like GR00T-N1.6 and Cosmos Policy achieves state-of-the-art results on RoboCasa and LIBERO, and improves real-world performance by 6.7%, enhancing embodied agent generalization.

Meng Cao, Haoran Tang, Haoze Zhao et al.

Recent advancements in video-based large language models (Video LLMs) have witnessed the emergence of diverse capabilities to reason and interpret dynamic visual content. Among them, gameplay videos stand out as a distinctive data source, often containing glitches that defy physics commonsense. This characteristic renders them an effective benchmark for assessing the under-explored capability of physical commonsense understanding in video LLMs. In this paper, we propose PhysGame as a pioneering benchmark to evaluate physical commonsense violations in gameplay videos. PhysGame comprises 880 videos associated with glitches spanning four fundamental domains (i.e., mechanics, kinematics, optics, and material properties) and across 12 distinct physical commonsense. Through extensively evaluating various state-ofthe-art video LLMs, our findings reveal that the performance of current open-source video LLMs significantly lags behind that of proprietary counterparts. To bridge this gap, we curate an instruction tuning dataset PhysInstruct with 140,057 question-answering pairs to facilitate physical commonsense learning. In addition, we also propose a preference optimization dataset PhysDPO with 34,358 training pairs, where the dis-preferred responses are generated conditioned on misleading titles (i.e., meta information hacking), fewer frames (i.e., temporal hacking) and lower spatial resolutions (i.e., spatial hacking). Based on the suite of datasets, we propose PhysVLM as a physical knowledge-enhanced video LLM. Extensive experiments on both physical-oriented benchmark PhysGame and general video understanding benchmarks demonstrate the state-ofthe-art performance of PhysVLM.

Zeyu Zhang, Danning Li, Ian Reid et al.

Energy-based predictive world models provide a powerful approach for multi-step visual planning by reasoning over latent energy landscapes rather than generating pixels. However, existing approaches face two major challenges: (i) their latent representations are typically learned in Euclidean space, neglecting the underlying geometric and hierarchical structure among states, and (ii) they struggle with long-horizon prediction, which leads to rapid degradation across extended rollouts. To address these challenges, we introduce GeoWorld, a geometric world model that preserves geometric structure and hierarchical relations through a Hyperbolic JEPA, which maps latent representations from Euclidean space onto hyperbolic manifolds. We further introduce Geometric Reinforcement Learning for energy-based optimization, enabling stable multi-step planning in hyperbolic latent space. Extensive experiments on CrossTask and COIN demonstrate around 3% SR improvement in 3-step planning and 2% SR improvement in 4-step planning compared to the state-of-the-art V-JEPA 2. Project website: https://steve-zeyu-zhang.github.io/GeoWorld.

Hu Wang, Congbo Ma, Ian Reid et al.

The advantage function is a central concept in RL that helps reduce variance in policy gradient estimates. Recently, for language modeling, Group Relative Policy Optimization (GRPO) was proposed to compute the advantage for each output by subtracting the mean reward, as the baseline, for all outputs in the group. However, it can lead to high variance when the reward advantage is inaccurately predicted. In this work, we propose Kalman Filter Enhanced Group Relative Policy Optimization (KRPO) model, by using lightweight Kalman filtering to dynamically estimate the latent reward baseline and uncertainty. This filtering technique replaces the naive group mean, enabling more adaptive advantage normalization. Our method does not require additional learned parameters over GRPO. This approach offers a simple yet effective way to incorporate multiple outputs of GRPO into advantage estimation, improving policy optimization in settings where highly dynamic reward signals are difficult to model for language models. Through the accuracies and rewards obtained from math question answering and reasoning, we show that using a more adaptive advantage estimation model, KRPO can improve the stability and performance of GRPO. The code is available at https://github.com/billhhh/KRPO_LLMs_RL.

Stefan Podgorski, Sourav Garg, Mehdi Hosseinzadeh et al.

Visual navigation in robotics traditionally relies on globally-consistent 3D maps or learned controllers, which can be computationally expensive and difficult to generalize across diverse environments. In this work, we present a novel RGB-only, object-level topometric navigation pipeline that enables zero-shot, long-horizon robot navigation without requiring 3D maps or pre-trained controllers. Our approach integrates global topological path planning with local metric trajectory control, allowing the robot to navigate towards object-level sub-goals while avoiding obstacles. We address key limitations of previous methods by continuously predicting local trajectory using monocular depth and traversability estimation, and incorporating an auto-switching mechanism that falls back to a baseline controller when necessary. The system operates using foundational models, ensuring open-set applicability without the need for domain-specific fine-tuning. We demonstrate the effectiveness of our method in both simulated environments and real-world tests, highlighting its robustness and deployability. Our approach outperforms existing state-of-the-art methods, offering a more adaptable and effective solution for visual navigation in open-set environments. The source code is made publicly available: https://github.com/podgorki/TANGO.

Hu Wang, Congbo Ma, Ibrahim Almakky et al.

Model merging, particularly through weight averaging, has shown surprising effectiveness in saving computations and improving model performance without any additional training. However, the interpretability of why and how this technique works remains unclear. In this work, we reinterpret weight-averaged model merging through the lens of interpretability and provide empirical insights into the underlying mechanisms that govern its behavior. We approach the problem from three perspectives: (1) we analyze the learned weight structures and demonstrate that model weights encode structured representations that help explain the compatibility of weight averaging; (2) we compare averaging in weight space and feature space across diverse model architectures (CNNs and ViTs) and datasets, aiming to expose under which circumstances what combination paradigm will work more effectively; (3) we study the effect of parameter scaling on prediction stability, highlighting how weight averaging acts as a form of regularization that contributes to robustness. By framing these analyses in an interpretability context, our work contributes to a more transparent and systematic understanding of model merging for stakeholders interested in the safety and reliability of untrained model combination methods. The code is available at https://github.com/billhhh/Rethink-Merge.

Meng Cao, Haoran Tang, Haoze Zhao et al.

Understanding the physical world, including object dynamics, material properties, and causal interactions, remains a core challenge in artificial intelligence. Although recent multi-modal large language models (MLLMs) have demonstrated impressive general reasoning capabilities, they still fall short of achieving human-level understanding of physical principles. Existing datasets for physical reasoning either rely on real-world videos, which incur high annotation costs, or on synthetic simulations, which suffer from limited realism and diversity. In this paper, we propose a novel paradigm that leverages glitches in gameplay videos, referring to visual anomalies that violate predefined physical laws, as a rich and scalable supervision source for physical world understanding. We introduce PhysGame, an meta information guided instruction-tuning dataset containing 140,057 glitch-centric question-answer pairs across five physical domains and sixteen fine-grained categories. To ensure data accuracy, we design a prompting strategy that utilizes gameplay metadata such as titles and descriptions to guide high-quality QA generation. Complementing PhysGame, we construct GameBench, an expert-annotated benchmark with 880 glitch-identified gameplay videos designed to evaluate physical reasoning capabilities. Extensive experiments show that PhysGame significantly enhances both Game2Real transferability, improving the real world physical reasoning performance of Qwen2.5VL by 2.5% on PhysBench, and Game2General transferability, yielding a 1.9% gain on the MVBench benchmark. Moreover, PhysGame-tuned models achieve a 3.7% absolute improvement on GameBench, demonstrating enhanced robustness in detecting physical implausibilities. These results indicate that learning from gameplay anomalies offers a scalable and effective pathway toward advancing physical world understanding in multimodal intelligence.

Boying Li, Cheng Zhang, Weirong Chen et al.

Camera calibration is a fundamental prerequisite for reliable geometric perception, yet classical approaches rely on controlled acquisition setups that are impractical for in-the-wild imagery. Recent learning-based methods have shown promising results for single-view calibration, but inherently neglect geometric consistency across multiple views. We introduce CalibAnyView, a unified formulation that supports an arbitrary number of input views ($N \geq 1$) by explicitly modeling cross-view geometric consistency. To facilitate this, we construct a large-scale multi-view video dataset covering diverse real-world scenarios, including multiple camera models, dynamic scenes, realistic motion trajectories, and heterogeneous lens distortions. Building on this dataset, we develop a multi-view transformer that predicts dense perspective fields, which are further integrated into a geometric optimization framework to jointly estimate camera intrinsics and gravity direction. Extensive experiments demonstrate that CalibAnyView consistently outperforms state-of-the-art methods, achieves strong robustness under single-view settings, and further improves with multi-view inference, providing a reliable foundation for downstream tasks such as 3D reconstruction and robotic perception in the wild.

Hu Wang, Salma Hassan, Yuyuan Liu et al.

In multi-modal learning, some modalities are more influential than others, and their absence can have a significant impact on classification/segmentation accuracy. Addressing this challenge, we propose a novel approach called Meta-learned Modality-weighted Knowledge Distillation (MetaKD), which enables multi-modal models to maintain high accuracy even when key modalities are missing. MetaKD adaptively estimates the importance weight of each modality through a meta-learning process. These learned importance weights guide a pairwise modality-weighted knowledge distillation process, allowing high-importance modalities to transfer knowledge to lower-importance ones, resulting in robust performance despite missing inputs. Unlike previous methods in the field, which are often task-specific and require significant modifications, our approach is designed to work in multiple tasks (e.g., segmentation and classification) with minimal adaptation. Experimental results on five prevalent datasets, including three Brain Tumor Segmentation datasets (BraTS2018, BraTS2019 and BraTS2020), the Alzheimer's Disease Neuroimaging Initiative (ADNI) classification dataset and the Audiovision-MNIST classification dataset, demonstrate the proposed model is able to outperform the compared models by a large margin. The code is available at https://github.com/billhhh/MetaKD.

Meng Cao, Haokun Lin, Haoyuan Li et al.

Spatial reasoning, the ability to understand and interpret the 3D structure of the world, is a critical yet underdeveloped capability in Multimodal Large Language Models (MLLMs). Current methods predominantly rely on verbal descriptive tuning, which suffers from visual illiteracy, i.e., they learn spatial concepts through textual symbols alone, devoid of connection to their visual manifestations. To bridge this gap, this paper introduces MILO, an Implicit spatIaL wOrld modeling paradigm that simulates human-like spatial imagination. MILO integrates a visual generator to provide geometry-aware feedback, thereby implicitly grounding the MLLM's symbolic reasoning in perceptual experience. Complementing this paradigm, we propose RePE (Relative Positional Encoding), a novel encoding scheme that captures relative camera-pose transformations, offering superior performance over absolute coordinate systems. To support the training, we construct GeoGen, a large-scale Geometry-aware Generative dataset with approximately 2,241 videos and 67,827 observation-action-outcome triplets. Experiments demonstrate that our approach significantly enhances spatial reasoning capabilities across multiple baselines and benchmarks, offering a more holistic understanding of 3D space.

Meng Cao, Xingyu Li, Xue Liu et al.

Despite advancements in Multi-modal Large Language Models (MLLMs) for scene understanding, their performance on complex spatial reasoning tasks requiring mental simulation remains significantly limited. Current methods often rely on passive observation of spatial data, failing to internalize an active mental imagery process. To bridge this gap, we propose SpatialDreamer, a reinforcement learning framework that enables spatial reasoning through a closedloop process of active exploration, visual imagination via a world model, and evidence-grounded reasoning. To address the lack of fine-grained reward supervision in longhorizontal reasoning tasks, we propose Geometric Policy Optimization (GeoPO), which introduces tree-structured sampling and step-level reward estimation with geometric consistency constraints. Extensive experiments demonstrate that SpatialDreamer delivers highly competitive results across multiple challenging benchmarks, signifying a critical advancement in human-like active spatial mental simulation for MLLMs.

Raphael Lafargue, Yassir Bendou, Bastien Pasdeloup et al.

When training data is scarce, it is common to make use of a feature extractor that has been pre-trained on a large base dataset, either by fine-tuning its parameters on the ``target'' dataset or by directly adopting its representation as features for a simple classifier. Fine-tuning is ineffective for few-shot learning, since the target dataset contains only a handful of examples. However, directly adopting the features without fine-tuning relies on the base and target distributions being similar enough that these features achieve separability and generalization. This paper investigates whether better features for the target dataset can be obtained by training on fewer base classes, seeking to identify a more useful base dataset for a given task.We consider cross-domain few-shot image classification in eight different domains from Meta-Dataset and entertain multiple real-world settings (domain-informed, task-informed and uninformed) where progressively less detail is known about the target task. To our knowledge, this is the first demonstration that fine-tuning on a subset of carefully selected base classes can significantly improve few-shot learning. Our contributions are simple and intuitive methods that can be implemented in any few-shot solution. We also give insights into the conditions in which these solutions are likely to provide a boost in accuracy. We release the code to reproduce all experiments from this paper on GitHub. https://github.com/RafLaf/Few-and-Fewer.git

Junlin Han, Pengfei Fang, Weihao Li et al.

We present You Only Cut Once (YOCO) for performing data augmentations. YOCO cuts one image into two pieces and performs data augmentations individually within each piece. Applying YOCO improves the diversity of the augmentation per sample and encourages neural networks to recognize objects from partial information. YOCO enjoys the properties of parameter-free, easy usage, and boosting almost all augmentations for free. Thorough experiments are conducted to evaluate its effectiveness. We first demonstrate that YOCO can be seamlessly applied to varying data augmentations, neural network architectures, and brings performance gains on CIFAR and ImageNet classification tasks, sometimes surpassing conventional image-level augmentation by large margins. Moreover, we show YOCO benefits contrastive pre-training toward a more powerful representation that can be better transferred to multiple downstream tasks. Finally, we study a number of variants of YOCO and empirically analyze the performance for respective settings. Code is available at GitHub.

Filipe R. Cordeiro, Vasileios Belagiannis, Ian Reid et al.

The most competitive noisy label learning methods rely on an unsupervised classification of clean and noisy samples, where samples classified as noisy are re-labelled and "MixMatched" with the clean samples. These methods have two issues in large noise rate problems: 1) the noisy set is more likely to contain hard samples that are in-correctly re-labelled, and 2) the number of samples produced by MixMatch tends to be reduced because it is constrained by the small clean set size. In this paper, we introduce the learning algorithm PropMix to handle the issues above. PropMix filters out hard noisy samples, with the goal of increasing the likelihood of correctly re-labelling the easy noisy samples. Also, PropMix places clean and re-labelled easy noisy samples in a training set that is augmented with MixUp, removing the clean set size constraint and including a large proportion of correctly re-labelled easy noisy samples. We also include self-supervised pre-training to improve robustness to high noisy label scenarios. Our experiments show that PropMix has state-of-the-art (SOTA) results on CIFAR-10/-100(with symmetric, asymmetric and semantic label noise), Red Mini-ImageNet (from the Controlled Noisy Web Labels), Clothing1M and WebVision. In severe label noise bench-marks, our results are substantially better than other methods. The code is available athttps://github.com/filipe-research/PropMix.

Filipe R. Cordeiro, Ragav Sachdeva, Vasileios Belagiannis et al.

Deep neural network models are robust to a limited amount of label noise, but their ability to memorise noisy labels in high noise rate problems is still an open issue. The most competitive noisy-label learning algorithms rely on a 2-stage process comprising an unsupervised learning to classify training samples as clean or noisy, followed by a semi-supervised learning that minimises the empirical vicinal risk (EVR) using a labelled set formed by samples classified as clean, and an unlabelled set with samples classified as noisy. In this paper, we hypothesise that the generalisation of such 2-stage noisy-label learning methods depends on the precision of the unsupervised classifier and the size of the training set to minimise the EVR. We empirically validate these two hypotheses and propose the new 2-stage noisy-label training algorithm LongReMix. We test LongReMix on the noisy-label benchmarks CIFAR-10, CIFAR-100, WebVision, Clothing1M, and Food101-N. The results show that our LongReMix generalises better than competing approaches, particularly in high label noise problems. Furthermore, our approach achieves state-of-the-art performance in most datasets. The code is available at https://github.com/filipe-research/LongReMix.

Huangying Zhan, Chamara Saroj Weerasekera, Jia-Wang Bian et al.

Multi-view geometry-based methods dominate the last few decades in monocular Visual Odometry for their superior performance, while they have been vulnerable to dynamic and low-texture scenes. More importantly, monocular methods suffer from scale-drift issue, i.e., errors accumulate over time. Recent studies show that deep neural networks can learn scene depths and relative camera in a self-supervised manner without acquiring ground truth labels. More surprisingly, they show that the well-trained networks enable scale-consistent predictions over long videos, while the accuracy is still inferior to traditional methods because of ignoring geometric information. Building on top of recent progress in computer vision, we design a simple yet robust VO system by integrating multi-view geometry and deep learning on Depth and optical Flow, namely DF-VO. In this work, a) we propose a method to carefully sample high-quality correspondences from deep flows and recover accurate camera poses with a geometric module; b) we address the scale-drift issue by aligning geometrically triangulated depths to the scale-consistent deep depths, where the dynamic scenes are taken into account. Comprehensive ablation studies show the effectiveness of the proposed method, and extensive evaluation results show the state-of-the-art performance of our system, e.g., Ours (1.652%) v.s. ORB-SLAM (3.247%}) in terms of translation error in KITTI Odometry benchmark. Source code is publicly available at: \href{https://github.com/Huangying-Zhan/DF-VO}{DF-VO}.

Ragav Sachdeva, Filipe R. Cordeiro, Vasileios Belagiannis et al.

The efficacy of deep learning depends on large-scale data sets that have been carefully curated with reliable data acquisition and annotation processes. However, acquiring such large-scale data sets with precise annotations is very expensive and time-consuming, and the cheap alternatives often yield data sets that have noisy labels. The field has addressed this problem by focusing on training models under two types of label noise: 1) closed-set noise, where some training samples are incorrectly annotated to a training label other than their known true class; and 2) open-set noise, where the training set includes samples that possess a true class that is (strictly) not contained in the set of known training labels. In this work, we study a new variant of the noisy label problem that combines the open-set and closed-set noisy labels, and introduce a benchmark evaluation to assess the performance of training algorithms under this setup. We argue that such problem is more general and better reflects the noisy label scenarios in practice. Furthermore, we propose a novel algorithm, called EvidentialMix, that addresses this problem and compare its performance with the state-of-the-art methods for both closed-set and open-set noise on the proposed benchmark. Our results show that our method produces superior classification results and better feature representations than previous state-of-the-art methods. The code is available at https://github.com/ragavsachdeva/EvidentialMix.

Haokui Zhang, Yu Liu, Bei Fang et al.

Hyperspectral image(HSI) classification has been improved with convolutional neural network(CNN) in very recent years. Being different from the RGB datasets, different HSI datasets are generally captured by various remote sensors and have different spectral configurations. Moreover, each HSI dataset only contains very limited training samples and thus it is prone to overfitting when using deep CNNs. In this paper, we first deliver a 3D asymmetric inception network, AINet, to overcome the overfitting problem. With the emphasis on spectral signatures over spatial contexts of HSI data, AINet can convey and classify the features effectively. In addition, the proposed data fusion transfer learning strategy is beneficial in boosting the classification performance. Extensive experiments show that the proposed approach beat all of the state-of-art methods on several HSI benchmarks, including Pavia University, Indian Pines and Kennedy Space Center(KSC). Code can be found at: https://github.com/UniLauX/AINet.

Yu Liu, Jie Li, Xia Yuan et al.

Semantic Scene Completion (SSC) refers to the task of inferring the 3D semantic segmentation of a scene while simultaneously completing the 3D shapes. We propose PALNet, a novel hybrid network for SSC based on single depth. PALNet utilizes a two-stream network to extract both 2D and 3D features from multi-stages using fine-grained depth information to efficiently captures the context, as well as the geometric cues of the scene. Current methods for SSC treat all parts of the scene equally causing unnecessary attention to the interior of objects. To address this problem, we propose Position Aware Loss(PA-Loss) which is position importance aware while training the network. Specifically, PA-Loss considers Local Geometric Anisotropy to determine the importance of different positions within the scene. It is beneficial for recovering key details like the boundaries of objects and the corners of the scene. Comprehensive experiments on two benchmark datasets demonstrate the effectiveness of the proposed method and its superior performance. Models and Video demo can be found at: https://github.com/UniLauX/PALNet.

Vladimir Nekrasov, Chunhua Shen, Ian Reid

Automatic search of neural architectures for various vision and natural language tasks is becoming a prominent tool as it allows to discover high-performing structures on any dataset of interest. Nevertheless, on more difficult domains, such as dense per-pixel classification, current automatic approaches are limited in their scope - due to their strong reliance on existing image classifiers they tend to search only for a handful of additional layers with discovered architectures still containing a large number of parameters. In contrast, in this work we propose a novel solution able to find light-weight and accurate segmentation architectures starting from only few blocks of a pre-trained classification network. To this end, we progressively build up a methodology that relies on templates of sets of operations, predicts which template and how many times should be applied at each step, while also generating the connectivity structure and downsampling factors. All these decisions are being made by a recurrent neural network that is rewarded based on the score of the emitted architecture on the holdout set and trained using reinforcement learning. One discovered architecture achieves 63.2% mean IoU on CamVid and 67.8% on CityScapes having only 270K parameters. Pre-trained models and the search code are available at https://github.com/DrSleep/nas-segm-pytorch.

Anh-Dzung Doan, Yasir Latif, Tat-Jun Chin et al.

A major focus of current research on place recognition is visual localization for autonomous driving. In this scenario, as cameras will be operating continuously, it is realistic to expect videos as an input to visual localization algorithms, as opposed to the single-image querying approach used in other visual localization works. In this paper, we show that exploiting temporal continuity in the testing sequence significantly improves visual localization - qualitatively and quantitatively. Although intuitive, this idea has not been fully explored in recent works. To this end, we propose two filtering approaches to exploit the temporal smoothness of image sequences: i) filtering on discrete domain with Hidden Markov Model, and ii) filtering on continuous domain with Monte Carlo-based visual localization. Our approaches rely on local features with an encoding technique to represent an image as a single vector. The experimental results on synthetic and real datasets show that our proposed methods achieve better results than state of the art (i.e., deep learning-based pose regression approaches) for the task on visual localization under significant appearance change. Our synthetic dataset and source code are made publicly available at https://sites.google.com/view/g2d-software/home and https://github.com/dadung/Visual-Localization-Filtering.

Vladimir Nekrasov, Hao Chen, Chunhua Shen et al.

Automated design of neural network architectures tailored for a specific task is an extremely promising, albeit inherently difficult, avenue to explore. While most results in this domain have been achieved on image classification and language modelling problems, here we concentrate on dense per-pixel tasks, in particular, semantic image segmentation using fully convolutional networks. In contrast to the aforementioned areas, the design choices of a fully convolutional network require several changes, ranging from the sort of operations that need to be used---e.g., dilated convolutions---to a solving of a more difficult optimisation problem. In this work, we are particularly interested in searching for high-performance compact segmentation architectures, able to run in real-time using limited resources. To achieve that, we intentionally over-parameterise the architecture during the training time via a set of auxiliary cells that provide an intermediate supervisory signal and can be omitted during the evaluation phase. The design of the auxiliary cell is emitted by a controller, a neural network with the fixed structure trained using reinforcement learning. More crucially, we demonstrate how to efficiently search for these architectures within limited time and computational budgets. In particular, we rely on a progressive strategy that terminates non-promising architectures from being further trained, and on Polyak averaging coupled with knowledge distillation to speed-up the convergence. Quantitatively, in 8 GPU-days our approach discovers a set of architectures performing on-par with state-of-the-art among compact models on the semantic segmentation, pose estimation and depth prediction tasks. Code will be made available here: https://github.com/drsleep/nas-segm-pytorch

Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera et al.

Despite learning based methods showing promising results in single view depth estimation and visual odometry, most existing approaches treat the tasks in a supervised manner. Recent approaches to single view depth estimation explore the possibility of learning without full supervision via minimizing photometric error. In this paper, we explore the use of stereo sequences for learning depth and visual odometry. The use of stereo sequences enables the use of both spatial (between left-right pairs) and temporal (forward backward) photometric warp error, and constrains the scene depth and camera motion to be in a common, real-world scale. At test time our framework is able to estimate single view depth and two-view odometry from a monocular sequence. We also show how we can improve on a standard photometric warp loss by considering a warp of deep features. We show through extensive experiments that: (i) jointly training for single view depth and visual odometry improves depth prediction because of the additional constraint imposed on depths and achieves competitive results for visual odometry; (ii) deep feature-based warping loss improves upon simple photometric warp loss for both single view depth estimation and visual odometry. Our method outperforms existing learning based methods on the KITTI driving dataset in both tasks. The source code is available at https://github.com/Huangying-Zhan/Depth-VO-Feat

Thanh-Toan Do, Anh Nguyen, Ian Reid

We propose AffordanceNet, a new deep learning approach to simultaneously detect multiple objects and their affordances from RGB images. Our AffordanceNet has two branches: an object detection branch to localize and classify the object, and an affordance detection branch to assign each pixel in the object to its most probable affordance label. The proposed framework employs three key components for effectively handling the multiclass problem in the affordance mask: a sequence of deconvolutional layers, a robust resizing strategy, and a multi-task loss function. The experimental results on the public datasets show that our AffordanceNet outperforms recent state-of-the-art methods by a fair margin, while its end-to-end architecture allows the inference at the speed of 150ms per image. This makes our AffordanceNet well suitable for real-time robotic applications. Furthermore, we demonstrate the effectiveness of AffordanceNet in different testing environments and in real robotic applications. The source code is available at https://github.com/nqanh/affordance-net

Vijay Kumar B G, Gustavo Carneiro, Ian Reid

Recent innovations in training deep convolutional neural network (ConvNet) models have motivated the design of new methods to automatically learn local image descriptors. The latest deep ConvNets proposed for this task consist of a siamese network that is trained by penalising misclassification of pairs of local image patches. Current results from machine learning show that replacing this siamese by a triplet network can improve the classification accuracy in several problems, but this has yet to be demonstrated for local image descriptor learning. Moreover, current siamese and triplet networks have been trained with stochastic gradient descent that computes the gradient from individual pairs or triplets of local image patches, which can make them prone to overfitting. In this paper, we first propose the use of triplet networks for the problem of local image descriptor learning. Furthermore, we also propose the use of a global loss that minimises the overall classification error in the training set, which can improve the generalisation capability of the model. Using the UBC benchmark dataset for comparing local image descriptors, we show that the triplet network produces a more accurate embedding than the siamese network in terms of the UBC dataset errors. Moreover, we also demonstrate that a combination of the triplet and global losses produces the best embedding in the field, using this triplet network. Finally, we also show that the use of the central-surround siamese network trained with the global loss produces the best result of the field on the UBC dataset. Pre-trained models are available online at https://github.com/vijaykbg/deep-patchmatch

Zeyu Zhang, Akide Liu, Ian Reid et al.

Human motion generation stands as a significant pursuit in generative computer vision, while achieving long-sequence and efficient motion generation remains challenging. Recent advancements in state space models (SSMs), notably Mamba, have showcased considerable promise in long sequence modeling with an efficient hardware-aware design, which appears to be a promising direction to build motion generation model upon it. Nevertheless, adapting SSMs to motion generation faces hurdles since the lack of a specialized design architecture to model motion sequence. To address these challenges, we propose Motion Mamba, a simple and efficient approach that presents the pioneering motion generation model utilized SSMs. Specifically, we design a Hierarchical Temporal Mamba (HTM) block to process temporal data by ensemble varying numbers of isolated SSM modules across a symmetric U-Net architecture aimed at preserving motion consistency between frames. We also design a Bidirectional Spatial Mamba (BSM) block to bidirectionally process latent poses, to enhance accurate motion generation within a temporal frame. Our proposed method achieves up to 50% FID improvement and up to 4 times faster on the HumanML3D and KIT-ML datasets compared to the previous best diffusion-based method, which demonstrates strong capabilities of high-quality long sequence motion modeling and real-time human motion generation. See project website https://steve-zeyu-zhang.github.io/MotionMamba/

Jing Wu, Jia-Wang Bian, Xinghui Li et al. · bytedance, oxford

We propose GaussCtrl, a text-driven method to edit a 3D scene reconstructed by the 3D Gaussian Splatting (3DGS). Our method first renders a collection of images by using the 3DGS and edits them by using a pre-trained 2D diffusion model (ControlNet) based on the input prompt, which is then used to optimise the 3D model. Our key contribution is multi-view consistent editing, which enables editing all images together instead of iteratively editing one image while updating the 3D model as in previous works. It leads to faster editing as well as higher visual quality. This is achieved by the two terms: (a) depth-conditioned editing that enforces geometric consistency across multi-view images by leveraging naturally consistent depth maps. (b) attention-based latent code alignment that unifies the appearance of edited images by conditioning their editing to several reference views through self and cross-view attention between images' latent representations. Experiments demonstrate that our method achieves faster editing and better visual results than previous state-of-the-art methods.

Mehdi Hosseinzadeh, Shin-Fang Chng, Yi Xu et al.

3D Gaussians have recently emerged as an effective scene representation for real-time splatting and accurate novel-view synthesis, motivating several works to adapt multi-view structure prediction networks to regress per-pixel 3D Gaussians from images. However, most prior work extends these networks to predict additional Gaussian parameters -- orientation, scale, opacity, and appearance -- while relying almost exclusively on view-synthesis supervision. We show that a view-synthesis loss alone is insufficient to recover geometrically meaningful splats in this setting. We analyze and address the ambiguities of learning 3D Gaussian splats under self-supervision for pose-free generalizable splatting, and introduce G3Splat, which enforces geometric priors to obtain geometrically consistent 3D scene representations. Trained on RE10K, our approach achieves state-of-the-art performance in (i) geometrically consistent reconstruction, (ii) relative pose estimation, and (iii) novel-view synthesis. We further demonstrate strong zero-shot generalization on ScanNet, substantially outperforming prior work in both geometry recovery and relative pose estimation. Code and pretrained models are released on our project page (https://m80hz.github.io/g3splat/).

Jiajun Deng, Tianyu He, Li Jiang et al.

Current 3D Large Multimodal Models (3D LMMs) have shown tremendous potential in 3D-vision-based dialogue and reasoning. However, how to further enhance 3D LMMs to achieve fine-grained scene understanding and facilitate flexible human-agent interaction remains a challenging problem. In this work, we introduce 3D-LLaVA, a simple yet highly powerful 3D LMM designed to act as an intelligent assistant in comprehending, reasoning, and interacting with the 3D world. Unlike existing top-performing methods that rely on complicated pipelines-such as offline multi-view feature extraction or additional task-specific heads-3D-LLaVA adopts a minimalist design with integrated architecture and only takes point clouds as input. At the core of 3D-LLaVA is a new Omni Superpoint Transformer (OST), which integrates three functionalities: (1) a visual feature selector that converts and selects visual tokens, (2) a visual prompt encoder that embeds interactive visual prompts into the visual token space, and (3) a referring mask decoder that produces 3D masks based on text description. This versatile OST is empowered by the hybrid pretraining to obtain perception priors and leveraged as the visual connector that bridges the 3D data to the LLM. After performing unified instruction tuning, our 3D-LLaVA reports impressive results on various benchmarks.

Kehan Chen, Dong An, Yan Huang et al.

We address the task of Vision-Language Navigation in Continuous Environments (VLN-CE) under the zero-shot setting. Zero-shot VLN-CE is particularly challenging due to the absence of expert demonstrations for training and minimal environment structural prior to guide navigation. To confront these challenges, we propose a Constraint-Aware Navigator (CA-Nav), which reframes zero-shot VLN-CE as a sequential, constraint-aware sub-instruction completion process. CA-Nav continuously translates sub-instructions into navigation plans using two core modules: the Constraint-Aware Sub-instruction Manager (CSM) and the Constraint-Aware Value Mapper (CVM). CSM defines the completion criteria for decomposed sub-instructions as constraints and tracks navigation progress by switching sub-instructions in a constraint-aware manner. CVM, guided by CSM's constraints, generates a value map on the fly and refines it using superpixel clustering to improve navigation stability. CA-Nav achieves the state-of-the-art performance on two VLN-CE benchmarks, surpassing the previous best method by 12 percent and 13 percent in Success Rate on the validation unseen splits of R2R-CE and RxR-CE, respectively. Moreover, CA-Nav demonstrates its effectiveness in real-world robot deployments across various indoor scenes and instructions.