The increase in available computing power and the Deep Learning revolution have allowed the exploration of new topics and frontiers in Artificial Intelligence research. A new field called Embodied Artificial Intelligence, which places at the intersection of Computer Vision, Robotics, and Decision Making, has been gaining importance during the last few years, as it aims to foster the development of smart autonomous robots and their deployment in society. The recent availability of large collections of 3D models for photorealistic robotic simulation has allowed faster and safe training of learning-based agents for millions of frames and a careful evaluation of their behavior before deploying the models on real robotic platforms. These intelligent agents are intended to perform a certain task in a possibly unknown environment. To this end, during the training in simulation, the agents learn to perform continuous interactions with the surroundings, such as gathering information from the environment, encoding and extracting useful cues for the task, and performing actions towards the final goal; where every action of the agent influences the interactions. This dissertation follows the complete creation process of embodied agents for indoor environments, from their concept to their implementation and deployment. We aim to contribute to research in Embodied AI and autonomous agents, in order to foster future work in this field. We present a detailed analysis of the procedure behind implementing an intelligent embodied agent, comprehending a thorough description of the current state-of-the-art in literature, technical explanations of the proposed methods, and accurate experimental studies on relevant robotic tasks.
Estimating the relative pose between two cameras is a fundamental step in many applications such as Structure-from-Motion. The common approach to relative pose estimation is to apply a minimal solver inside a RANSAC loop. Highly efficient solvers exist for pinhole cameras. Yet, (nearly) all cameras exhibit radial distortion. Not modeling radial distortion leads to (significantly) worse results. However, minimal radial distortion solvers are significantly more complex than pinhole solvers, both in terms of run-time and implementation efforts. This paper compares radial distortion solvers with two simple-to-implement approaches that do not use minimal radial distortion solvers: The first approach combines an efficient pinhole solver with sampled radial undistortion parameters, where the sampled parameters are used for undistortion prior to applying the pinhole solver. The second approach uses a state-of-the-art neural network to estimate the distortion parameters rather than sampling them from a set of potential values. Extensive experiments on multiple datasets, and different camera setups, show that complex minimal radial distortion solvers are not necessary in practice. We discuss under which conditions a simple sampling of radial undistortion parameters is preferable over calibrating cameras using a learning-based prior approach. Code and newly created benchmark for relative pose estimation under radial distortion are available at https://github.com/kocurvik/rdnet.
Recent advancements in large language models have demonstrated how chain-of-thought (CoT) and reinforcement learning (RL) can improve performance. However, applying such reasoning strategies to the visual generation domain remains largely unexplored. In this paper, we present T2I-R1, a novel reasoning-enhanced text-to-image generation model, powered by RL with a bi-level CoT reasoning process. Specifically, we identify two levels of CoT that can be utilized to enhance different stages of generation: (1) the semantic-level CoT for high-level planning of the prompt and (2) the token-level CoT for low-level pixel processing during patch-by-patch generation. To better coordinate these two levels of CoT, we introduce BiCoT-GRPO with an ensemble of generation rewards, which seamlessly optimizes both generation CoTs within the same training step. By applying our reasoning strategies to the baseline model, Janus-Pro, we achieve superior performance with 13% improvement on T2I-CompBench and 19% improvement on the WISE benchmark, even surpassing the state-of-the-art model FLUX.1. Code is available at: https://github.com/CaraJ7/T2I-R1
We introduce RegionFocus, a visual test-time scaling approach for Vision Language Model Agents. Understanding webpages is challenging due to the visual complexity of GUI images and the large number of interface elements, making accurate action selection difficult. Our approach dynamically zooms in on relevant regions, reducing background clutter and improving grounding accuracy. To support this process, we propose an image-as-map mechanism that visualizes key landmarks at each step, providing a transparent action record and enables the agent to effectively choose among action candidates. Even with a simple region selection strategy, we observe significant performance gains of 28+\% on Screenspot-pro and 24+\% on WebVoyager benchmarks on top of two state-of-the-art open vision language model agents, UI-TARS and Qwen2.5-VL, highlighting the effectiveness of visual test-time scaling in interactive settings. We achieve a new state-of-the-art grounding performance of 61.6\% on the ScreenSpot-Pro benchmark by applying RegionFocus to a Qwen2.5-VL-72B model. Our code will be released publicly at https://github.com/tiangeluo/RegionFocus.
Multimodal LLMs are turning their focus to video benchmarks, however most video benchmarks only provide outcome supervision, with no intermediate or interpretable reasoning steps. This makes it challenging to assess if models are truly able to combine perceptual and temporal information to reason about videos, or simply get the correct answer by chance or by exploiting linguistic biases. To remedy this, we provide a new video reasoning dataset called MINERVA for modern multimodal models. Each question in the dataset comes with 5 answer choices, as well as detailed, hand-crafted reasoning traces. Our dataset is multimodal, diverse in terms of video domain and length, and consists of complex multi-step questions. Extensive benchmarking shows that our dataset provides a challenge for frontier open-source and proprietary models. We perform fine-grained error analysis to identify common failure modes across various models, and create a taxonomy of reasoning errors. We use this to explore both human and LLM-as-a-judge methods for scoring video reasoning traces, and find that failure modes are primarily related to temporal localization, followed by visual perception errors, as opposed to logical or completeness errors. The dataset, along with questions, answer candidates and reasoning traces will be publicly available under https://github.com/google-deepmind/neptune?tab=readme-ov-file\#minerva.
Detecting and tracking objects is a crucial component of any autonomous navigation method. For the past decades, object detection has yielded promising results using neural networks on various datasets. While many methods focus on performance metrics, few projects focus on improving the robustness of these detection and tracking pipelines, notably to sensor failures. In this paper we attempt to address this issue by creating a realistic synthetic data augmentation pipeline for camera-radar Autonomous Vehicle (AV) datasets. Our goal is to accurately simulate sensor failures and data deterioration due to real-world interferences. We also present our results of a baseline lightweight Noise Recognition neural network trained and tested on our augmented dataset, reaching an overall recognition accuracy of 54.4\% on 11 categories across 10086 images and 2145 radar point-clouds.
Locating specific segments within an instructional video is an efficient way to acquire guiding knowledge. Generally, the task of obtaining video segments for both verbal explanations and visual demonstrations is known as visual answer localization (VAL). However, users often need multiple interactions to obtain answers that align with their expectations when using the system. During these interactions, humans deepen their understanding of the video content by asking themselves questions, thereby accurately identifying the location. Therefore, we propose a new task, named In-VAL, to simulate the multiple interactions between humans and videos in the procedure of obtaining visual answers. The In-VAL task requires interactively addressing several semantic gap issues, including 1) the ambiguity of user intent in the input questions, 2) the incompleteness of language in video subtitles, and 3) the fragmentation of content in video segments. To address these issues, we propose Ask2Loc, a framework for resolving In-VAL by asking questions. It includes three key modules: 1) a chatting module to refine initial questions and uncover clear intentions, 2) a rewriting module to generate fluent language and create complete descriptions, and 3) a searching module to broaden local context and provide integrated content. We conduct extensive experiments on three reconstructed In-VAL datasets. Compared to traditional end-to-end and two-stage methods, our proposed Ask2Loc can improve performance by up to 14.91 (mIoU) on the In-VAL task. Our code and datasets can be accessed at https://github.com/changzong/Ask2Loc.
Prompt engineering for large language models is challenging, as even small prompt perturbations or model changes can significantly impact the generated output texts. Existing evaluation methods of LLM outputs, either automated metrics or human evaluation, have limitations, such as providing limited insights or being labor-intensive. We propose Spotlight, a new approach that combines both automation and human analysis. Based on data mining techniques, we automatically distinguish between random (decoding) variations and systematic differences in language model outputs. This process provides token patterns that describe the systematic differences and guide the user in manually analyzing the effects of their prompts and changes in models efficiently. We create three benchmarks to quantitatively test the reliability of token pattern extraction methods and demonstrate that our approach provides new insights into established prompt data. From a human-centric perspective, through demonstration studies and a user study, we show that our token pattern approach helps users understand the systematic differences of language model outputs. We are further able to discover relevant differences caused by prompt and model changes (e.g. related to gender or culture), thus supporting the prompt engineering process and human-centric model behavior research.
Diffusion models approximate the denoising distribution as a Gaussian and predict its mean, whereas flow matching models reparameterize the Gaussian mean as flow velocity. However, they underperform in few-step sampling due to discretization error and tend to produce over-saturated colors under classifier-free guidance (CFG). To address these limitations, we propose a novel Gaussian mixture flow matching (GMFlow) model: instead of predicting the mean, GMFlow predicts dynamic Gaussian mixture (GM) parameters to capture a multi-modal flow velocity distribution, which can be learned with a KL divergence loss. We demonstrate that GMFlow generalizes previous diffusion and flow matching models where a single Gaussian is learned with an $L_2$ denoising loss. For inference, we derive GM-SDE/ODE solvers that leverage analytic denoising distributions and velocity fields for precise few-step sampling. Furthermore, we introduce a novel probabilistic guidance scheme that mitigates the over-saturation issues of CFG and improves image generation quality. Extensive experiments demonstrate that GMFlow consistently outperforms flow matching baselines in generation quality, achieving a Precision of 0.942 with only 6 sampling steps on ImageNet 256$\times$256.
Humans have the tendency to discover and explore. This natural tendency is reflected in data from streaming platforms as the amount of previously unknown content accessed by users. Additionally, in domains such as that of music streaming there is evidence that recommending novel content improves users' experience with the platform. Therefore, understanding users' discovery patterns, such as the amount to which and the way users access previously unknown content, is a topic of relevance for both the scientific community and the streaming industry, particularly the music one. Previous works studied how music consumption differs for users of different traits and looked at diversity, novelty, and consistency over time of users' music preferences. However, very little is known about how users discover and explore previously unknown music, and how this behavior differs for users of varying discovery needs. In this paper we bridge this gap by analyzing data from a survey answered by users of the major music streaming platform Deezer in combination with their streaming data. We first address questions regarding whether users who declare a higher interest in unfamiliar music listen to more diverse music, have more stable music preferences over time, and explore more music within a same time window, compared to those who declare a lower interest. We then investigate which type of music tracks users choose to listen to when they explore unfamiliar music, identifying clear patterns of popularity and genre representativeness that vary for users of different discovery needs. Our findings open up possibilities to infer users' interest in unfamiliar music from streaming data as well as possibilities to develop recommender systems that guide users in exploring music in a more natural way.
Web-based educational videos offer flexible learning opportunities and are becoming increasingly popular. However, improving user engagement and knowledge retention remains a challenge. Automatically generated questions can activate learners and support their knowledge acquisition. Further, they can help teachers and learners assess their understanding. While large language and vision-language models have been employed in various tasks, their application to question generation for educational videos remains underexplored. In this paper, we investigate the capabilities of current vision-language models for generating learning-oriented questions for educational video content. We assess (1) out-of-the-box models' performance; (2) fine-tuning effects on content-specific question generation; (3) the impact of different video modalities on question quality; and (4) in a qualitative study, question relevance, answerability, and difficulty levels of generated questions. Our findings delineate the capabilities of current vision-language models, highlighting the need for fine-tuning and addressing challenges in question diversity and relevance. We identify requirements for future multimodal datasets and outline promising research directions.
Pancreatic cancer, characterized by its notable prevalence and mortality rates, demands accurate lesion delineation for effective diagnosis and therapeutic interventions. The generalizability of extant methods is frequently compromised due to the pronounced variability in imaging and the heterogeneous characteristics of pancreatic lesions, which may mimic normal tissues and exhibit significant inter-patient variability. Thus, we propose a generalization framework that synergizes pixel-level classification and regression tasks, to accurately delineate lesions and improve model stability. This framework not only seeks to align segmentation contours with actual lesions but also uses regression to elucidate spatial relationships between diseased and normal tissues, thereby improving tumor localization and morphological characterization. Enhanced by the reciprocal transformation of task outputs, our approach integrates additional regression supervision within the segmentation context, bolstering the model's generalization ability from a dual-task perspective. Besides, dual self-supervised learning in feature spaces and output spaces augments the model's representational capability and stability across different imaging views. Experiments on 594 samples composed of three datasets with significant imaging differences demonstrate that our generalized pancreas segmentation results comparable to mainstream in-domain validation performance (Dice: 84.07%). More importantly, it successfully improves the results of the highly challenging cross-lesion generalized pancreatic cancer segmentation task by 9.51%. Thus, our model constitutes a resilient and efficient foundational technological support for pancreatic disease management and wider medical applications. The codes will be released at https://github.com/SJTUBME-QianLab/Dual-Task-Seg.
Remote sensing enables a wide range of critical applications such as land cover and land use mapping, crop yield prediction, and environmental monitoring. Advances in satellite technology have expanded remote sensing datasets, yet high-performance segmentation models remain dependent on extensive labeled data, challenged by annotation scarcity and variability across sensors, illumination, and geography. Domain adaptation offers a promising solution to improve model generalization. This paper introduces a domain generalization approach to leveraging emerging geospatial foundation models by combining soft-alignment pseudo-labeling with source-to-target generative pre-training. We further provide new mathematical insights into MAE-based generative learning for domain-invariant feature learning. Experiments with hyperspectral and multispectral remote sensing datasets confirm our method's effectiveness in enhancing adaptability and segmentation.
Synthetic video generation has gained significant attention for its realism and broad applications, but remains prone to violations of common sense and physical laws. This highlights the need for reliable abnormality detectors that understand such principles and are robust to hallucinations. To address this, we introduce VideoHallu, a benchmark of over 3,000 video QA pairs built from synthetic videos generated by models like Veo2, Sora, and Kling, paired with expert-crafted counterintuitive QA to evaluate the critical thinking abilities of Multi-modal Large Language Models (MLLMs) on abnormalities that are perceptually obvious to humans but often hallucinated due to language priors. VideoHallu evaluates MLLMs' abnormality detection abilities with examples across alignment, consistency, commonsense, and physics. We benchmark SOTA MLLMs, including GPT-4o, Gemini-2.5-Pro, Qwen2.5-VL, Video-R1, and VideoChat-R1. We observe that these models perform well on many real-world benchmarks like MVBench and MovieChat, but still struggle with basic physics-based and commonsense reasoning in synthetic videos. We further show that post-training with Group Relative Policy Optimization (GRPO), using curriculum learning on datasets combining video QA with counterintuitive commonsense and physics reasoning over real and synthetic videos, improves MLLMs' abnormality detection and critical thinking, demonstrating the value of targeted training for improving their understanding of commonsense and physical laws.
Online multi-object tracking has been recently dominated by tracking-by-detection (TbD) methods, where recent advances rely on increasingly sophisticated heuristics for tracklet representation, feature fusion, and multi-stage matching. The key strength of TbD lies in its modular design, enabling the integration of specialized off-the-shelf models like motion predictors and re-identification. However, the extensive usage of human-crafted rules for temporal associations makes these methods inherently limited in their ability to capture the complex interplay between various tracking cues. In this work, we introduce CAMEL, a novel association module for Context-Aware Multi-Cue ExpLoitation, that learns resilient association strategies directly from data, breaking free from hand-crafted heuristics while maintaining TbD's valuable modularity. At its core, CAMEL employs two transformer-based modules and relies on a novel association-centric training scheme to effectively model the complex interactions between tracked targets and their various association cues. Unlike end-to-end detection-by-tracking approaches, our method remains lightweight and fast to train while being able to leverage external off-the-shelf models. Our proposed online tracking pipeline, CAMELTrack, achieves state-of-the-art performance on multiple tracking benchmarks. Our code is available at https://github.com/TrackingLaboratory/CAMELTrack.
Recent advances in audio-visual learning have shown promising results in learning representations across modalities. However, most approaches rely on global audio representations that fail to capture fine-grained temporal correspondences with visual frames. Additionally, existing methods often struggle with conflicting optimization objectives when trying to jointly learn reconstruction and cross-modal alignment. In this work, we propose CAV-MAE Sync as a simple yet effective extension of the original CAV-MAE framework for self-supervised audio-visual learning. We address three key challenges: First, we tackle the granularity mismatch between modalities by treating audio as a temporal sequence aligned with video frames, rather than using global representations. Second, we resolve conflicting optimization goals by separating contrastive and reconstruction objectives through dedicated global tokens. Third, we improve spatial localization by introducing learnable register tokens that reduce semantic load on patch tokens. We evaluate the proposed approach on AudioSet, VGG Sound, and the ADE20K Sound dataset on zero-shot retrieval, classification and localization tasks demonstrating state-of-the-art performance and outperforming more complex architectures.
The increasing accessibility of remotely sensed data and the potential of such data to inform large-scale decision-making has driven the development of deep learning models for many Earth Observation tasks. Traditionally, such models must be trained on large datasets. However, the common assumption that broadly larger datasets lead to better outcomes tends to overlook the complexities of the data distribution, the potential for introducing biases and noise, and the computational resources required for processing and storing vast datasets. Therefore, effective solutions should consider both the quantity and quality of data. In this paper, we propose six novel core-set selection methods for selecting important subsets of samples from remote sensing image segmentation datasets that rely on imagery only, labels only, and a combination of each. We benchmark these approaches against a random-selection baseline on three commonly used land cover classification datasets: DFC2022, Vaihingen, and Potsdam. In each of the datasets, we demonstrate that training on a subset of samples outperforms the random baseline, and some approaches outperform training on all available data. This result shows the importance and potential of data-centric learning for the remote sensing domain. The code is available at https://github.com/keillernogueira/data-centric-rs-classification/.
Underwater image enhancement (UIE) is a critical preprocessing step for marine vision applications, where wavelength-dependent attenuation causes severe content degradation and color distortion. While recent state space models like Mamba show potential for long-range dependency modeling, their unfolding operations and fixed scan paths on 1D sequences fail to adapt to local object semantics and global relation modeling, limiting their efficacy in complex underwater environments. To address this, we enhance conventional Mamba with the sorting-based scanning mechanism that dynamically reorders scanning sequences based on statistical distribution of spatial correlation of all pixels. In this way, it encourages the network to prioritize the most informative components--structural and semantic features. Upon building this mechanism, we devise a Visually Self-adaptive State Block (VSSB) that harmonizes dynamic sorting of Mamba with input-dependent dynamic convolution, enabling coherent integration of global context and local relational cues. This exquisite design helps eliminate global focus bias, especially for widely distributed contents, which greatly weakens the statistical frequency. For robust feature extraction and refinement, we design a cross-feature bridge (CFB) to adaptively fuse multi-scale representations. These efforts compose the novel relation-driven Mamba framework for effective UIE (RD-UIE). Extensive experiments on underwater enhancement benchmarks demonstrate RD-UIE outperforms the state-of-the-art approach WMamba in both quantitative metrics and visual fidelity, averagely achieving 0.55 dB performance gain on the three benchmarks. Our code is available at https://github.com/kkoucy/RD-UIE/tree/main
High Dynamic Range Novel View Synthesis (HDR-NVS) aims to establish a 3D scene HDR model from Low Dynamic Range (LDR) imagery. Typically, multiple-exposure LDR images are employed to capture a wider range of brightness levels in a scene, as a single LDR image cannot represent both the brightest and darkest regions simultaneously. While effective, this multiple-exposure HDR-NVS approach has significant limitations, including susceptibility to motion artifacts (e.g., ghosting and blurring), high capture and storage costs. To overcome these challenges, we introduce, for the first time, the single-exposure HDR-NVS problem, where only single exposure LDR images are available during training. We further introduce a novel approach, Mono-HDR-3D, featuring two dedicated modules formulated by the LDR image formation principles, one for converting LDR colors to HDR counterparts, and the other for transforming HDR images to LDR format so that unsupervised learning is enabled in a closed loop. Designed as a meta-algorithm, our approach can be seamlessly integrated with existing NVS models. Extensive experiments show that Mono-HDR-3D significantly outperforms previous methods. Source code will be released.
Long video generation involves generating extended videos using models trained on short videos, suffering from distribution shifts due to varying frame counts. It necessitates the use of local information from the original short frames to enhance visual and motion quality, and global information from the entire long frames to ensure appearance consistency. Existing training-free methods struggle to effectively integrate the benefits of both, as appearance and motion in videos are closely coupled, leading to motion inconsistency and visual quality. In this paper, we reveal that global and local information can be precisely decoupled into consistent appearance and motion intensity information by applying Principal Component Analysis (PCA), allowing for refined complementary integration of global consistency and local quality. With this insight, we propose FreePCA, a training-free long video generation paradigm based on PCA that simultaneously achieves high consistency and quality. Concretely, we decouple consistent appearance and motion intensity features by measuring cosine similarity in the principal component space. Critically, we progressively integrate these features to preserve original quality and ensure smooth transitions, while further enhancing consistency by reusing the mean statistics of the initial noise. Experiments demonstrate that FreePCA can be applied to various video diffusion models without requiring training, leading to substantial improvements. Code is available at https://github.com/JosephTiTan/FreePCA.