Bharath Hariharan

Mark Hamilton, Zhoutong Zhang, Bharath Hariharan et al. · deepmind

Unsupervised semantic segmentation aims to discover and localize semantically meaningful categories within image corpora without any form of annotation. To solve this task, algorithms must produce features for every pixel that are both semantically meaningful and compact enough to form distinct clusters. Unlike previous works which achieve this with a single end-to-end framework, we propose to separate feature learning from cluster compactification. Empirically, we show that current unsupervised feature learning frameworks already generate dense features whose correlations are semantically consistent. This observation motivates us to design STEGO ($\textbf{S}$elf-supervised $\textbf{T}$ransformer with $\textbf{E}$nergy-based $\textbf{G}$raph $\textbf{O}$ptimization), a novel framework that distills unsupervised features into high-quality discrete semantic labels. At the core of STEGO is a novel contrastive loss function that encourages features to form compact clusters while preserving their relationships across the corpora. STEGO yields a significant improvement over the prior state of the art, on both the CocoStuff ($\textbf{+14 mIoU}$) and Cityscapes ($\textbf{+9 mIoU}$) semantic segmentation challenges.

Qianqian Wang, Yen-Yu Chang, Ruojin Cai et al. · deepmind

We present a new test-time optimization method for estimating dense and long-range motion from a video sequence. Prior optical flow or particle video tracking algorithms typically operate within limited temporal windows, struggling to track through occlusions and maintain global consistency of estimated motion trajectories. We propose a complete and globally consistent motion representation, dubbed OmniMotion, that allows for accurate, full-length motion estimation of every pixel in a video. OmniMotion represents a video using a quasi-3D canonical volume and performs pixel-wise tracking via bijections between local and canonical space. This representation allows us to ensure global consistency, track through occlusions, and model any combination of camera and object motion. Extensive evaluations on the TAP-Vid benchmark and real-world footage show that our approach outperforms prior state-of-the-art methods by a large margin both quantitatively and qualitatively. See our project page for more results: http://omnimotion.github.io/

Menglin Jia, Luming Tang, Bor-Chun Chen et al. · deepmind

The current modus operandi in adapting pre-trained models involves updating all the backbone parameters, ie, full fine-tuning. This paper introduces Visual Prompt Tuning (VPT) as an efficient and effective alternative to full fine-tuning for large-scale Transformer models in vision. Taking inspiration from recent advances in efficiently tuning large language models, VPT introduces only a small amount (less than 1% of model parameters) of trainable parameters in the input space while keeping the model backbone frozen. Via extensive experiments on a wide variety of downstream recognition tasks, we show that VPT achieves significant performance gains compared to other parameter efficient tuning protocols. Most importantly, VPT even outperforms full fine-tuning in many cases across model capacities and training data scales, while reducing per-task storage cost.

Luming Tang, Menglin Jia, Qianqian Wang et al. · deepmind

Finding correspondences between images is a fundamental problem in computer vision. In this paper, we show that correspondence emerges in image diffusion models without any explicit supervision. We propose a simple strategy to extract this implicit knowledge out of diffusion networks as image features, namely DIffusion FeaTures (DIFT), and use them to establish correspondences between real images. Without any additional fine-tuning or supervision on the task-specific data or annotations, DIFT is able to outperform both weakly-supervised methods and competitive off-the-shelf features in identifying semantic, geometric, and temporal correspondences. Particularly for semantic correspondence, DIFT from Stable Diffusion is able to outperform DINO and OpenCLIP by 19 and 14 accuracy points respectively on the challenging SPair-71k benchmark. It even outperforms the state-of-the-art supervised methods on 9 out of 18 categories while remaining on par for the overall performance. Project page: https://diffusionfeatures.github.io

Ruojin Cai, Joseph Tung, Qianqian Wang et al. · deepmind

We consider the visual disambiguation task of determining whether a pair of visually similar images depict the same or distinct 3D surfaces (e.g., the same or opposite sides of a symmetric building). Illusory image matches, where two images observe distinct but visually similar 3D surfaces, can be challenging for humans to differentiate, and can also lead 3D reconstruction algorithms to produce erroneous results. We propose a learning-based approach to visual disambiguation, formulating it as a binary classification task on image pairs. To that end, we introduce a new dataset for this problem, Doppelgangers, which includes image pairs of similar structures with ground truth labels. We also design a network architecture that takes the spatial distribution of local keypoints and matches as input, allowing for better reasoning about both local and global cues. Our evaluation shows that our method can distinguish illusory matches in difficult cases, and can be integrated into SfM pipelines to produce correct, disambiguated 3D reconstructions. See our project page for our code, datasets, and more results: http://doppelgangers-3d.github.io/.

Guandao Yang, Sagie Benaim, Varun Jampani et al.

Neural fields have emerged as a new paradigm for representing signals, thanks to their ability to do it compactly while being easy to optimize. In most applications, however, neural fields are treated like black boxes, which precludes many signal manipulation tasks. In this paper, we propose a new class of neural fields called polynomial neural fields (PNFs). The key advantage of a PNF is that it can represent a signal as a composition of a number of manipulable and interpretable components without losing the merits of neural fields representation. We develop a general theoretical framework to analyze and design PNFs. We use this framework to design Fourier PNFs, which match state-of-the-art performance in signal representation tasks that use neural fields. In addition, we empirically demonstrate that Fourier PNFs enable signal manipulation applications such as texture transfer and scale-space interpolation. Code is available at https://github.com/stevenygd/PNF.

Luming Tang, Nataniel Ruiz, Qinghao Chu et al. · deepmind

Recent advances in generative imagery have brought forth outpainting and inpainting models that can produce high-quality, plausible image content in unknown regions. However, the content these models hallucinate is necessarily inauthentic, since they are unaware of the true scene. In this work, we propose RealFill, a novel generative approach for image completion that fills in missing regions of an image with the content that should have been there. RealFill is a generative inpainting model that is personalized using only a few reference images of a scene. These reference images do not have to be aligned with the target image, and can be taken with drastically varying viewpoints, lighting conditions, camera apertures, or image styles. Once personalized, RealFill is able to complete a target image with visually compelling contents that are faithful to the original scene. We evaluate RealFill on a new image completion benchmark that covers a set of diverse and challenging scenarios, and find that it outperforms existing approaches by a large margin. Project page: https://realfill.github.io

Travis Zhang, Katie Luo, Cheng Perng Phoo et al.

The rapid development of 3D object detection systems for self-driving cars has significantly improved accuracy. However, these systems struggle to generalize across diverse driving environments, which can lead to safety-critical failures in detecting traffic participants. To address this, we propose a method that utilizes unlabeled repeated traversals of multiple locations to adapt object detectors to new driving environments. By incorporating statistics computed from repeated LiDAR scans, we guide the adaptation process effectively. Our approach enhances LiDAR-based detection models using spatial quantized historical features and introduces a lightweight regression head to leverage the statistics for feature regularization. Additionally, we leverage the statistics for a novel self-training process to stabilize the training. The framework is detector model-agnostic and experiments on real-world datasets demonstrate significant improvements, achieving up to a 20-point performance gain, especially in detecting pedestrians and distant objects. Code is available at https://github.com/zhangtravis/Hist-DA.

Gemmechu Hassena, Jonathan Moon, Ryan Fujii et al. · deepmind

Implicit neural fields have made remarkable progress in reconstructing 3D surfaces from multiple images; however, they encounter challenges when it comes to separating individual objects within a scene. Previous work has attempted to tackle this problem by introducing a framework to train separate signed distance fields (SDFs) simultaneously for each of N objects and using a regularization term to prevent objects from overlapping. However, all of these methods require segmentation masks to be provided, which are not always readily available. We introduce our method, ObjectCarver, to tackle the problem of object separation from just click input in a single view. Given posed multi-view images and a set of user-input clicks to prompt segmentation of the individual objects, our method decomposes the scene into separate objects and reconstructs a high-quality 3D surface for each one. We introduce a loss function that prevents floaters and avoids inappropriate carving-out due to occlusion. In addition, we introduce a novel scene initialization method that significantly speeds up the process while preserving geometric details compared to previous approaches. Despite requiring neither ground truth masks nor monocular cues, our method outperforms baselines both qualitatively and quantitatively. In addition, we introduce a new benchmark dataset for evaluation.

Carlos A. Diaz-Ruiz, Youya Xia, Yurong You et al.

Advances in perception for self-driving cars have accelerated in recent years due to the availability of large-scale datasets, typically collected at specific locations and under nice weather conditions. Yet, to achieve the high safety requirement, these perceptual systems must operate robustly under a wide variety of weather conditions including snow and rain. In this paper, we present a new dataset to enable robust autonomous driving via a novel data collection process - data is repeatedly recorded along a 15 km route under diverse scene (urban, highway, rural, campus), weather (snow, rain, sun), time (day/night), and traffic conditions (pedestrians, cyclists and cars). The dataset includes images and point clouds from cameras and LiDAR sensors, along with high-precision GPS/INS to establish correspondence across routes. The dataset includes road and object annotations using amodal masks to capture partial occlusions and 3D bounding boxes. We demonstrate the uniqueness of this dataset by analyzing the performance of baselines in amodal segmentation of road and objects, depth estimation, and 3D object detection. The repeated routes opens new research directions in object discovery, continual learning, and anomaly detection. Link to Ithaca365: https://ithaca365.mae.cornell.edu/

Samar Khanna, Bram Wallace, Kavita Bala et al.

Geographic variance in satellite imagery impacts the ability of machine learning models to generalise to new regions. In this paper, we model geographic generalisation in medium resolution Landsat-8 satellite imagery as a continuous domain adaptation problem, demonstrating how models generalise better with appropriate domain knowledge. We develop a dataset spatially distributed across the entire continental United States, providing macroscopic insight into the effects of geography on crop classification in multi-spectral and temporally distributed satellite imagery. Our method demonstrates improved generalisability from 1) passing geographically correlated climate variables along with the satellite data to a Transformer model and 2) regressing on the model features to reconstruct these domain variables. Combined, we provide a novel perspective on geographic generalisation in satellite imagery and a simple-yet-effective approach to leverage domain knowledge. Code is available at: \url{https://github.com/samar-khanna/cropmap}

Yurong You, Katie Z Luo, Xiangyu Chen et al.

Self-driving cars must detect vehicles, pedestrians, and other traffic participants accurately to operate safely. Small, far-away, or highly occluded objects are particularly challenging because there is limited information in the LiDAR point clouds for detecting them. To address this challenge, we leverage valuable information from the past: in particular, data collected in past traversals of the same scene. We posit that these past data, which are typically discarded, provide rich contextual information for disambiguating the above-mentioned challenging cases. To this end, we propose a novel, end-to-end trainable Hindsight framework to extract this contextual information from past traversals and store it in an easy-to-query data structure, which can then be leveraged to aid future 3D object detection of the same scene. We show that this framework is compatible with most modern 3D detection architectures and can substantially improve their average precision on multiple autonomous driving datasets, most notably by more than 300% on the challenging cases.

Youya Xia, Josephine Monica, Wei-Lun Chao et al.

A self-driving car must be able to reliably handle adverse weather conditions (e.g., snowy) to operate safely. In this paper, we investigate the idea of turning sensor inputs (i.e., images) captured in an adverse condition into a benign one (i.e., sunny), upon which the downstream tasks (e.g., semantic segmentation) can attain high accuracy. Prior work primarily formulates this as an unpaired image-to-image translation problem due to the lack of paired images captured under the exact same camera poses and semantic layouts. While perfectly-aligned images are not available, one can easily obtain coarsely-paired images. For instance, many people drive the same routes daily in both good and adverse weather; thus, images captured at close-by GPS locations can form a pair. Though data from repeated traversals are unlikely to capture the same foreground objects, we posit that they provide rich contextual information to supervise the image translation model. To this end, we propose a novel training objective leveraging coarsely-aligned image pairs. We show that our coarsely-aligned training scheme leads to a better image translation quality and improved downstream tasks, such as semantic segmentation, monocular depth estimation, and visual localization.

Davis Wertheimer, Luming Tang, Bharath Hariharan · deepmind

Few-shot recognition involves training an image classifier to distinguish novel concepts at test time using few examples (shot). Existing approaches generally assume that the shot number at test time is known in advance. This is not realistic, and the performance of a popular and foundational method has been shown to suffer when train and test shots do not match. We conduct a systematic empirical study of this phenomenon. In line with prior work, we find that shot sensitivity is broadly present across metric-based few-shot learners, but in contrast to prior work, larger neural architectures provide a degree of built-in robustness to varying test shot. More importantly, a simple, previously known but greatly overlooked class of approaches based on cosine distance consistently and greatly improves robustness to shot variation, by removing sensitivity to sample noise. We derive cosine alternatives to popular and recent few-shot classifiers, broadening their applicability to realistic settings. These cosine models consistently improve shot-robustness, outperform prior shot-robust state of the art, and provide competitive accuracy on a range of benchmarks and architectures, including notable gains in the very-low-shot regime.

Katie Z Luo, Zhenzhen Liu, Xiangyu Chen et al.

Recent advances in machine learning have shown that Reinforcement Learning from Human Feedback (RLHF) can improve machine learning models and align them with human preferences. Although very successful for Large Language Models (LLMs), these advancements have not had a comparable impact in research for autonomous vehicles -- where alignment with human expectations can be imperative. In this paper, we propose to adapt similar RL-based methods to unsupervised object discovery, i.e. learning to detect objects from LiDAR points without any training labels. Instead of labels, we use simple heuristics to mimic human feedback. More explicitly, we combine multiple heuristics into a simple reward function that positively correlates its score with bounding box accuracy, i.e., boxes containing objects are scored higher than those without. We start from the detector's own predictions to explore the space and reinforce boxes with high rewards through gradient updates. Empirically, we demonstrate that our approach is not only more accurate, but also orders of magnitudes faster to train compared to prior works on object discovery.

Yurong You, Cheng Perng Phoo, Katie Z Luo et al.

For a self-driving car to operate reliably, its perceptual system must generalize to the end-user's environment -- ideally without additional annotation efforts. One potential solution is to leverage unlabeled data (e.g., unlabeled LiDAR point clouds) collected from the end-users' environments (i.e. target domain) to adapt the system to the difference between training and testing environments. While extensive research has been done on such an unsupervised domain adaptation problem, one fundamental problem lingers: there is no reliable signal in the target domain to supervise the adaptation process. To overcome this issue we observe that it is easy to collect unsupervised data from multiple traversals of repeated routes. While different from conventional unsupervised domain adaptation, this assumption is extremely realistic since many drivers share the same roads. We show that this simple additional assumption is sufficient to obtain a potent signal that allows us to perform iterative self-training of 3D object detectors on the target domain. Concretely, we generate pseudo-labels with the out-of-domain detector but reduce false positives by removing detections of supposedly mobile objects that are persistent across traversals. Further, we reduce false negatives by encouraging predictions in regions that are not persistent. We experiment with our approach on two large-scale driving datasets and show remarkable improvement in 3D object detection of cars, pedestrians, and cyclists, bringing us a step closer to generalizable autonomous driving.

Tai-Yu Pan, Chenyang Ma, Tianle Chen et al.

Accurate 3D object detection and understanding for self-driving cars heavily relies on LiDAR point clouds, necessitating large amounts of labeled data to train. In this work, we introduce an innovative pre-training approach, Grounded Point Colorization (GPC), to bridge the gap between data and labels by teaching the model to colorize LiDAR point clouds, equipping it with valuable semantic cues. To tackle challenges arising from color variations and selection bias, we incorporate color as "context" by providing ground-truth colors as hints during colorization. Experimental results on the KITTI and Waymo datasets demonstrate GPC's remarkable effectiveness. Even with limited labeled data, GPC significantly improves fine-tuning performance; notably, on just 20% of the KITTI dataset, GPC outperforms training from scratch with the entire dataset. In sum, we introduce a fresh perspective on pre-training for 3D object detection, aligning the objective with the model's intended role and ultimately advancing the accuracy and efficiency of 3D object detection for autonomous vehicles.

Kenneth Borup, Cheng Perng Phoo, Bharath Hariharan

We address the challenge of getting efficient yet accurate recognition systems with limited labels. While recognition models improve with model size and amount of data, many specialized applications of computer vision have severe resource constraints both during training and inference. Transfer learning is an effective solution for training with few labels, however often at the expense of a computationally costly fine-tuning of large base models. We propose to mitigate this unpleasant trade-off between compute and accuracy via semi-supervised cross-domain distillation from a set of diverse source models. Initially, we show how to use task similarity metrics to select a single suitable source model to distill from, and that a good selection process is imperative for good downstream performance of a target model. We dub this approach DistillNearest. Though effective, DistillNearest assumes a single source model matches the target task, which is not always the case. To alleviate this, we propose a weighted multi-source distillation method to distill multiple source models trained on different domains weighted by their relevance for the target task into a single efficient model (named DistillWeighted). Our methods need no access to source data, and merely need features and pseudo-labels of the source models. When the goal is accurate recognition under computational constraints, both DistillNearest and DistillWeighted approaches outperform both transfer learning from strong ImageNet initializations as well as state-of-the-art semi-supervised techniques such as FixMatch. Averaged over 8 diverse target tasks our multi-source method outperforms the baselines by 5.6%-points and 4.5%-points, respectively.

Chia-Hsiang Kao, Bharath Hariharan

Despite its widespread use in neural networks, error backpropagation has faced criticism for its lack of biological plausibility, suffering from issues such as the backward locking problem and the weight transport problem. These limitations have motivated researchers to explore more biologically plausible learning algorithms that could potentially shed light on how biological neural systems adapt and learn. Inspired by the counter-current exchange mechanisms observed in biological systems, we propose counter-current learning (CCL), a biologically plausible framework for credit assignment in neural networks. This framework employs a feedforward network to process input data and a feedback network to process targets, with each network enhancing the other through anti-parallel signal propagation. By leveraging the more informative signals from the bottom layer of the feedback network to guide the updates of the top layer of the feedforward network and vice versa, CCL enables the simultaneous transformation of source inputs to target outputs and the dynamic mutual influence of these transformations. Experimental results on MNIST, FashionMNIST, CIFAR10, and CIFAR100 datasets using multi-layer perceptrons and convolutional neural networks demonstrate that CCL achieves comparable performance to other biologically plausible algorithms while offering a more biologically realistic learning mechanism. Furthermore, we showcase the applicability of our approach to an autoencoder task, underscoring its potential for unsupervised representation learning. Our work presents a direction for biologically inspired and plausible learning algorithms, offering an alternative mechanism of learning and adaptation in neural networks.

Yihong Sun, Bharath Hariharan

Unsupervised monocular depth estimation techniques have demonstrated encouraging results but typically assume that the scene is static. These techniques suffer when trained on dynamical scenes, where apparent object motion can equally be explained by hypothesizing the object's independent motion, or by altering its depth. This ambiguity causes depth estimators to predict erroneous depth for moving objects. To resolve this issue, we introduce Dynamo-Depth, an unifying approach that disambiguates dynamical motion by jointly learning monocular depth, 3D independent flow field, and motion segmentation from unlabeled monocular videos. Specifically, we offer our key insight that a good initial estimation of motion segmentation is sufficient for jointly learning depth and independent motion despite the fundamental underlying ambiguity. Our proposed method achieves state-of-the-art performance on monocular depth estimation on Waymo Open and nuScenes Dataset with significant improvement in the depth of moving objects. Code and additional results are available at https://dynamo-depth.github.io.

Weikai Huang, Jieyu Zhang, Sijun Li et al.

Understanding objects in 3D from a single image is a cornerstone of spatial intelligence. A key step toward this goal is monocular 3D object detection--recovering the extent, location, and orientation of objects from an input RGB image. To be practical in the open world, such a detector must generalize beyond closed-set categories, support diverse prompt modalities, and leverage geometric cues when available. Progress is hampered by two bottlenecks: existing methods are designed for a single prompt type and lack a mechanism to incorporate additional geometric cues, and current 3D datasets cover only narrow categories in controlled environments, limiting open-world transfer. In this work we address both gaps. First, we introduce WildDet3D, a unified geometry-aware architecture that natively accepts text, point, and box prompts and can incorporate auxiliary depth signals at inference time. Second, we present WildDet3D-Data, the largest open 3D detection dataset to date, constructed by generating candidate 3D boxes from existing 2D annotations and retaining only human-verified ones, yielding over 1M images across 13.5K categories in diverse real-world scenes. WildDet3D establishes a new state-of-the-art across multiple benchmarks and settings. In the open-world setting, it achieves 22.6/24.8 AP3D on our newly introduced WildDet3D-Bench with text and box prompts. On Omni3D, it reaches 34.2/36.4 AP3D with text and box prompts, respectively. In zero-shot evaluation, it achieves 40.3/48.9 ODS on Argoverse 2 and ScanNet. Notably, incorporating depth cues at inference time yields substantial additional gains (+20.7 AP on average across settings).

Zilu Li, Guandao Yang, Qingqing Zhao et al.

This paper presents a method to leverage arbitrary neural network architecture for control variates. Control variates are crucial in reducing the variance of Monte Carlo integration, but they hinge on finding a function that both correlates with the integrand and has a known analytical integral. Traditional approaches rely on heuristics to choose this function, which might not be expressive enough to correlate well with the integrand. Recent research alleviates this issue by modeling the integrands with a learnable parametric model, such as a neural network. However, the challenge remains in creating an expressive parametric model with a known analytical integral. This paper proposes a novel approach to construct learnable parametric control variates functions from arbitrary neural network architectures. Instead of using a network to approximate the integrand directly, we employ the network to approximate the anti-derivative of the integrand. This allows us to use automatic differentiation to create a function whose integration can be constructed by the antiderivative network. We apply our method to solve partial differential equations using the Walk-on-sphere algorithm. Our results indicate that this approach is unbiased and uses various network architectures to achieve lower variance than other control variate methods.

Wenxuan Peng, Bharath Hariharan, Hadar Averbuch-Elor

Despite recent progress, text-to-image models still struggle to generate semantically diverse and compositionally accurate multi-person interaction scenes, often collapsing to repetitive layouts, stereotypical poses, and poorly grounded interactions. In this work, we bridge this gap by introducing a dual pose-image representation that brings person-centric structural priors into pretrained diffusion transformers. Our model jointly predicts a 2D pose visualization image and its corresponding RGB image, enabling structure and appearance to co-evolve during learning. At its core, a cross-modal alignment scheme binds text, pose, and image representations, ensuring consistent grounding across modalities. Furthermore, we design an iterative scene construction scheme, progressively generating complex multi-human interactions while effectively decomposing the overall generation complexity. Extensive experiments demonstrate that our method substantially improves prompt alignment and scene diversity in multi-person image generation.

Zhen Xu, Jinsu Yoo, Cristian Bautista et al.

Building robust 3D perception for self-driving still relies heavily on large-scale data collection and manual annotation, yet this paradigm becomes impractical as deployment expands across diverse cities and regions. Meanwhile, modern cities are increasingly instrumented with roadside units (RSUs), static sensors deployed along roads and at intersections to monitor traffic. This raises a natural question: can the city itself help train the vehicle? We propose infrastructure-taught, label-free 3D perception, a paradigm in which RSUs act as stationary, unsupervised teachers for ego vehicles. Leveraging their fixed viewpoints and repeated observations, RSUs learn local 3D detectors from unlabeled data and broadcast predictions to passing vehicles, which are aggregated as pseudo-label supervision for training a standalone ego detector. The resulting model requires no infrastructure or communication at test time. We instantiate this idea as a fully label-free three-stage pipeline and conduct a concept-and-feasibility study in a CARLA-based multi-agent environment. With CenterPoint, our pipeline achieves 82.3% AP for detecting vehicles, compared to a fully supervised ego upper bound of 94.4%. We further systematically analyze each stage, evaluate its scalability, and demonstrate complementarity with existing ego-centric label-free methods. Together, these results suggest that city infrastructure itself can potentially provide a scalable supervisory signal for autonomous vehicles, positioning infrastructure-taught learning as a promising orthogonal paradigm for reducing annotation cost in 3D perception.

Chia-Hsiang Kao, Cong Phuoc Huynh, Chien-Yi Wang et al.

Inferring rigid-body physical states and properties from monocular videos is a fundamental step toward physics-based perception and simulation. Existing approaches assume specific underlying physical systems, object types, and camera poses, making them unable to generalize to complex real-world settings. We introduce $Δ$YNAMICS, a vision-language framework that uses language as a unified representation of rigid-body dynamics. Instead of directly predicting parameters, $Δ$YNAMICS generates scene configurations in a structured text format for physics simulation. We enhance the model's generalization by integrating natural language motion reasoning and leveraging optical flow as a semantic-agnostic input. On the CLEVRER dataset, $Δ$YNAMICS achieves a segmentation IoU of 0.30, a 7x improvement over leading VLMs (InternVL3-8B, Qwen2.5-VL-7B and Claude-4-Sonnet). Additionally, test-time sampling and evolutionary search further boost performance by 27% and 120% in segmentation IoU, respectively. Finally, we demonstrate strong transfer to a new dataset of 235 real-world rigid-body videos, highlighting the potential of language-driven physics inference for bridging perception and simulation.

Aditya Chetan, Eric Cai, Peeyush Kushwaha et al.

The emergence of Large Vision-Language Models (LVLMs) has significantly advanced video understanding capabilities. However, existing benchmarks focus predominantly on coarse-grained tasks such as action segmentation, classification, captioning, and retrieval. Furthermore, these benchmarks often rely on entities that can be easily identified verbally, like household objects, animals, human subjects, etc., limiting their applicability to complex, in-the-wild video scenarios. But, many applications such as furniture assembly, cooking, etc., require step-by-step fine-grained spatio-temporal understanding of the video, which is not sufficiently evaluated in current benchmarks. To address this gap, we introduce Flat-Pack Bench, a novel benchmark centered on furniture assembly tasks. Our benchmark evaluates LVLMs on nuanced tasks, including temporal ordering of assembly actions, temporal localization of assembly state, understanding part mating, and tracking, using multiple-choice questions paired with visual prompts highlighting relevant parts as references for fine-grained questions. Our experiments reveal that state-of-the-art LVLMs struggle significantly with fine-grained spatio-temporal reasoning, highlighting their limitations in effectively leveraging temporal information from videos, limited tracking ability, and understanding of spatial interactions like physical contact.

Xinyu Yang, Haozheng Yu, Yihong Sun et al.

Interactive video segmentation often requires many user interventions for robust performance in challenging scenarios (e.g., occlusions, object separations, camouflage, etc.). Yet, even state-of-the-art models like SAM2 use corrections only for immediate fixes without learning from this feedback, leading to inefficient, repetitive user effort. To address this, we introduce Live Interactive Training (LIT), a novel framework for prompt-based visual systems where models also learn online from human corrections at inference time. Our primary instantiation, LIT-LoRA, implements this by continually updating a lightweight LoRA module on-the-fly. When a user provides a correction, this module is rapidly trained on that feedback, allowing the vision system to improve performance on subsequent frames of the same video. Leveraging the core principles of LIT, our LIT-LoRA implementation achieves an average 18-34% reduction in total corrections on challenging video segmentation benchmarks, with a negligible training overhead of ~0.5s per correction. We further demonstrate its generality by successfully adapting it to other segmentation models and extending it to CLIP-based fine-grained image classification. Our work highlights the promise of live adaptation to transform interactive tools and significantly reduce redundant human effort in complex visual tasks. Project: https://youngxinyu1802.github.io/projects/LIT/.

Gene Chou, Wenqi Xian, Guandao Yang et al. · deepmind

A versatile video depth estimation model should (1) be accurate and consistent across frames, (2) produce high-resolution depth maps, and (3) support real-time streaming. We propose FlashDepth, a method that satisfies all three requirements, performing depth estimation on a 2044x1148 streaming video at 24 FPS. We show that, with careful modifications to pretrained single-image depth models, these capabilities are enabled with relatively little data and training. We evaluate our approach across multiple unseen datasets against state-of-the-art depth models, and find that ours outperforms them in terms of boundary sharpness and speed by a significant margin, while maintaining competitive accuracy. We hope our model will enable various applications that require high-resolution depth, such as video editing, and online decision-making, such as robotics. We release all code and model weights at https://github.com/Eyeline-Research/FlashDepth

Yihong Sun, Bharath Hariharan

Embodied agents must detect and localize objects of interest, e.g. traffic participants for self-driving cars. Supervision in the form of bounding boxes for this task is extremely expensive. As such, prior work has looked at unsupervised instance detection and segmentation, but in the absence of annotated boxes, it is unclear how pixels must be grouped into objects and which objects are of interest. This results in over-/under-segmentation and irrelevant objects. Inspired by human visual system and practical applications, we posit that the key missing cue for unsupervised detection is motion: objects of interest are typically mobile objects that frequently move and their motions can specify separate instances. In this paper, we propose MOD-UV, a Mobile Object Detector learned from Unlabeled Videos only. We begin with instance pseudo-labels derived from motion segmentation, but introduce a novel training paradigm to progressively discover small objects and static-but-mobile objects that are missed by motion segmentation. As a result, though only learned from unlabeled videos, MOD-UV can detect and segment mobile objects from a single static image. Empirically, we achieve state-of-the-art performance in unsupervised mobile object detection on Waymo Open, nuScenes, and KITTI Datasets without using any external data or supervised models. Code is available at https://github.com/YihongSun/MOD-UV.

Yihong Sun, Xinyu Yang, Jennifer J. Sun et al.

Real-world objects frequently undergo state transformations. From an apple being cut into pieces to a butterfly emerging from its cocoon, tracking through these changes is important for understanding real-world objects and dynamics. However, existing methods often lose track of the target object after transformation, due to significant changes in object appearance. To address this limitation, we introduce the task of Track Any State: tracking objects through transformations while detecting and describing state changes, accompanied by a new benchmark dataset, VOST-TAS. To tackle this problem, we present TubeletGraph, a zero-shot system that recovers missing objects after transformation and maps out how object states are evolving over time. TubeletGraph first identifies potentially overlooked tracks, and determines whether they should be integrated based on semantic and proximity priors. Then, it reasons about the added tracks and generates a state graph describing each observed transformation. TubeletGraph achieves state-of-the-art tracking performance under transformations, while demonstrating deeper understanding of object transformations and promising capabilities in temporal grounding and semantic reasoning for complex object transformations. Code, additional results, and the benchmark dataset are available at https://tubelet-graph.github.io.

Shaden Shaar, Bradon Thymes, Sirawut Chaixanien et al.

Understanding real-world videos such as movies requires integrating visual and dialogue cues to answer complex questions. Yet existing VideoQA benchmarks struggle to capture this multimodal reasoning and are largely not open-ended, given the difficulty of evaluating free-form answers. In this paper, we introduce a novel open-ended multi-modal VideoQA benchmark, MovieRecapsQA created using movie recap videos--a distinctive type of YouTube content that summarizes a film by presenting its key events through synchronized visual (recap video) and textual (recap summary) modalities. Using the recap summary, we generate $\approx 8.2$ K question-answer (QA) pairs (aligned with movie-subtitles) and provide the necessary "facts" needed to verify an answer in a reference-free manner. To our knowledge, this is the first open-ended VideoQA benchmark that supplies explicit textual context of the input (video and/or text); which we use for evaluation. Our benchmark provides videos of multiple lengths (i.e., recap-segments, movie-segments) and categorizations of questions (by modality and type) to enable fine-grained analysis. We evaluate the performance of seven state-of-the-art MLLMs using our benchmark and observe that: 1) visual-only questions remain the most challenging; 2) models default to textual inputs whenever available; 3) extracting factually accurate information from video content is still difficult for all models; and 4) proprietary and open-source models perform comparably on video-dependent questions.

Shreelekha Revankar, Utkarsh Mall, Cheng Perng Phoo et al.

Natural disasters cause devastating damage to communities and infrastructure every year. Effective disaster response is hampered by the difficulty of accessing affected areas during and after events. Remote sensing has allowed us to monitor natural disasters in a remote way. More recently there have been advances in computer vision and deep learning that help automate satellite imagery analysis, However, they remain limited by their narrow focus on specific disaster types, reliance on manual expert interpretation, and lack of datasets with sufficient temporal granularity or natural language annotations for tracking disaster progression. We present MONITRS, a novel multimodal dataset of more than 10,000 FEMA disaster events with temporal satellite imagery and natural language annotations from news articles, accompanied by geotagged locations, and question-answer pairs. We demonstrate that fine-tuning existing MLLMs on our dataset yields significant performance improvements for disaster monitoring tasks, establishing a new benchmark for machine learning-assisted disaster response systems. Code can be found at: https://github.com/ShreelekhaR/MONITRS

Jang Hyun Cho, Utkarsh Mall, Kavita Bala et al.

We present a new framework for semantic segmentation without annotations via clustering. Off-the-shelf clustering methods are limited to curated, single-label, and object-centric images yet real-world data are dominantly uncurated, multi-label, and scene-centric. We extend clustering from images to pixels and assign separate cluster membership to different instances within each image. However, solely relying on pixel-wise feature similarity fails to learn high-level semantic concepts and overfits to low-level visual cues. We propose a method to incorporate geometric consistency as an inductive bias to learn invariance and equivariance for photometric and geometric variations. With our novel learning objective, our framework can learn high-level semantic concepts. Our method, PiCIE (Pixel-level feature Clustering using Invariance and Equivariance), is the first method capable of segmenting both things and stuff categories without any hyperparameter tuning or task-specific pre-processing. Our method largely outperforms existing baselines on COCO and Cityscapes with +17.5 Acc. and +4.5 mIoU. We show that PiCIE gives a better initialization for standard supervised training. The code is available at https://github.com/janghyuncho/PiCIE.

Cheng Perng Phoo, Bharath Hariharan

Most few-shot learning techniques are pre-trained on a large, labeled "base dataset". In problem domains where such large labeled datasets are not available for pre-training (e.g., X-ray, satellite images), one must resort to pre-training in a different "source" problem domain (e.g., ImageNet), which can be very different from the desired target task. Traditional few-shot and transfer learning techniques fail in the presence of such extreme differences between the source and target tasks. In this paper, we present a simple and effective solution to tackle this extreme domain gap: self-training a source domain representation on unlabeled data from the target domain. We show that this improves one-shot performance on the target domain by 2.9 points on average on the challenging BSCD-FSL benchmark consisting of datasets from multiple domains. Our code is available at https://github.com/cpphoo/STARTUP.

Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor et al.

In this work, we propose a novel technique to generate shapes from point cloud data. A point cloud can be viewed as samples from a distribution of 3D points whose density is concentrated near the surface of the shape. Point cloud generation thus amounts to moving randomly sampled points to high-density areas. We generate point clouds by performing stochastic gradient ascent on an unnormalized probability density, thereby moving sampled points toward the high-likelihood regions. Our model directly predicts the gradient of the log density field and can be trained with a simple objective adapted from score-based generative models. We show that our method can reach state-of-the-art performance for point cloud auto-encoding and generation, while also allowing for extraction of a high-quality implicit surface. Code is available at https://github.com/RuojinCai/ShapeGF.

Divyansh Garg, Yan Wang, Bharath Hariharan et al.

Existing approaches to depth or disparity estimation output a distribution over a set of pre-defined discrete values. This leads to inaccurate results when the true depth or disparity does not match any of these values. The fact that this distribution is usually learned indirectly through a regression loss causes further problems in ambiguous regions around object boundaries. We address these issues using a new neural network architecture that is capable of outputting arbitrary depth values, and a new loss function that is derived from the Wasserstein distance between the true and the predicted distributions. We validate our approach on a variety of tasks, including stereo disparity and depth estimation, and the downstream 3D object detection. Our approach drastically reduces the error in ambiguous regions, especially around object boundaries that greatly affect the localization of objects in 3D, achieving the state-of-the-art in 3D object detection for autonomous driving. Our code will be available at https://github.com/Div99/W-Stereo-Disp.

Yan Wang, Xiangyu Chen, Yurong You et al.

In the domain of autonomous driving, deep learning has substantially improved the 3D object detection accuracy for LiDAR and stereo camera data alike. While deep networks are great at generalization, they are also notorious to over-fit to all kinds of spurious artifacts, such as brightness, car sizes and models, that may appear consistently throughout the data. In fact, most datasets for autonomous driving are collected within a narrow subset of cities within one country, typically under similar weather conditions. In this paper we consider the task of adapting 3D object detectors from one dataset to another. We observe that naively, this appears to be a very challenging task, resulting in drastic drops in accuracy levels. We provide extensive experiments to investigate the true adaptation challenges and arrive at a surprising conclusion: the primary adaptation hurdle to overcome are differences in car sizes across geographic areas. A simple correction based on the average car size yields a strong correction of the adaptation gap. Our proposed method is simple and easily incorporated into most 3D object detection frameworks. It provides a first baseline for 3D object detection adaptation across countries, and gives hope that the underlying problem may be more within grasp than one may have hoped to believe. Our code is available at https://github.com/cxy1997/3D_adapt_auto_driving.

Bram Wallace, Bharath Hariharan

Self-supervised representation learning has achieved impressive results in recent years, with experiments primarily coming on ImageNet or other similarly large internet imagery datasets. There has been little to no work with these methods on other smaller domains, such as satellite, textural, or biological imagery. We experiment with several popular methods on an unprecedented variety of domains. We discover, among other findings, that Rotation is by far the most semantically meaningful task, with much of the performance of Jigsaw and Instance Discrimination being attributable to the nature of their induced distribution rather than semantic understanding. Additionally, there are several areas, such as fine-grain classification, where all tasks underperform. We quantitatively and qualitatively diagnose the reasons for these failures and successes via novel experiments studying pretext generalization, random labelings, and implicit dimensionality. Code and models are available at https://github.com/BramSW/Extending_SSRL_Across_Domains/.

Rui Qian, Divyansh Garg, Yan Wang et al.

Reliable and accurate 3D object detection is a necessity for safe autonomous driving. Although LiDAR sensors can provide accurate 3D point cloud estimates of the environment, they are also prohibitively expensive for many settings. Recently, the introduction of pseudo-LiDAR (PL) has led to a drastic reduction in the accuracy gap between methods based on LiDAR sensors and those based on cheap stereo cameras. PL combines state-of-the-art deep neural networks for 3D depth estimation with those for 3D object detection by converting 2D depth map outputs to 3D point cloud inputs. However, so far these two networks have to be trained separately. In this paper, we introduce a new framework based on differentiable Change of Representation (CoR) modules that allow the entire PL pipeline to be trained end-to-end. The resulting framework is compatible with most state-of-the-art networks for both tasks and in combination with PointRCNN improves over PL consistently across all benchmarks -- yielding the highest entry on the KITTI image-based 3D object detection leaderboard at the time of submission. Our code will be made available at https://github.com/mileyan/pseudo-LiDAR_e2e.

Luming Tang, Davis Wertheimer, Bharath Hariharan

Few-shot, fine-grained classification requires a model to learn subtle, fine-grained distinctions between different classes (e.g., birds) based on a few images alone. This requires a remarkable degree of invariance to pose, articulation and background. A solution is to use pose-normalized representations: first localize semantic parts in each image, and then describe images by characterizing the appearance of each part. While such representations are out of favor for fully supervised classification, we show that they are extremely effective for few-shot fine-grained classification. With a minimal increase in model capacity, pose normalization improves accuracy between 10 and 20 percentage points for shallow and deep architectures, generalizes better to new domains, and is effective for multiple few-shot algorithms and network backbones. Code is available at https://github.com/Tsingularity/PoseNorm_Fewshot

Guandao Yang, Xun Huang, Zekun Hao et al.

As 3D point clouds become the representation of choice for multiple vision and graphics applications, the ability to synthesize or reconstruct high-resolution, high-fidelity point clouds becomes crucial. Despite the recent success of deep learning models in discriminative tasks of point clouds, generating point clouds remains challenging. This paper proposes a principled probabilistic framework to generate 3D point clouds by modeling them as a distribution of distributions. Specifically, we learn a two-level hierarchy of distributions where the first level is the distribution of shapes and the second level is the distribution of points given a shape. This formulation allows us to both sample shapes and sample an arbitrary number of points from a shape. Our generative model, named PointFlow, learns each level of the distribution with a continuous normalizing flow. The invertibility of normalizing flows enables the computation of the likelihood during training and allows us to train our model in the variational inference framework. Empirically, we demonstrate that PointFlow achieves state-of-the-art performance in point cloud generation. We additionally show that our model can faithfully reconstruct point clouds and learn useful representations in an unsupervised manner. The code will be available at https://github.com/stevenygd/PointFlow.

Yurong You, Yan Wang, Wei-Lun Chao et al.

Detecting objects such as cars and pedestrians in 3D plays an indispensable role in autonomous driving. Existing approaches largely rely on expensive LiDAR sensors for accurate depth information. While recently pseudo-LiDAR has been introduced as a promising alternative, at a much lower cost based solely on stereo images, there is still a notable performance gap. In this paper we provide substantial advances to the pseudo-LiDAR framework through improvements in stereo depth estimation. Concretely, we adapt the stereo network architecture and loss function to be more aligned with accurate depth estimation of faraway objects --- currently the primary weakness of pseudo-LiDAR. Further, we explore the idea to leverage cheaper but extremely sparse LiDAR sensors, which alone provide insufficient information for 3D detection, to de-bias our depth estimation. We propose a depth-propagation algorithm, guided by the initial depth estimates, to diffuse these few exact measurements across the entire depth map. We show on the KITTI object detection benchmark that our combined approach yields substantial improvements in depth estimation and stereo-based 3D object detection --- outperforming the previous state-of-the-art detection accuracy for faraway objects by 40%. Our code is available at https://github.com/mileyan/Pseudo_Lidar_V2.

Yan Wang, Wei-Lun Chao, Divyansh Garg et al.

3D object detection is an essential task in autonomous driving. Recent techniques excel with highly accurate detection rates, provided the 3D input data is obtained from precise but expensive LiDAR technology. Approaches based on cheaper monocular or stereo imagery data have, until now, resulted in drastically lower accuracies --- a gap that is commonly attributed to poor image-based depth estimation. However, in this paper we argue that it is not the quality of the data but its representation that accounts for the majority of the difference. Taking the inner workings of convolutional neural networks into consideration, we propose to convert image-based depth maps to pseudo-LiDAR representations --- essentially mimicking the LiDAR signal. With this representation we can apply different existing LiDAR-based detection algorithms. On the popular KITTI benchmark, our approach achieves impressive improvements over the existing state-of-the-art in image-based performance --- raising the detection accuracy of objects within the 30m range from the previous state-of-the-art of 22% to an unprecedented 74%. At the time of submission our algorithm holds the highest entry on the KITTI 3D object detection leaderboard for stereo-image-based approaches. Our code is publicly available at https://github.com/mileyan/pseudo_lidar.

Yan Wang, Lequn Wang, Yurong You et al.

Not all people are equally easy to identify: color statistics might be enough for some cases while others might require careful reasoning about high- and low-level details. However, prevailing person re-identification(re-ID) methods use one-size-fits-all high-level embeddings from deep convolutional networks for all cases. This might limit their accuracy on difficult examples or makes them needlessly expensive for the easy ones. To remedy this, we present a new person re-ID model that combines effective embeddings built on multiple convolutional network layers, trained with deep-supervision. On traditional re-ID benchmarks, our method improves substantially over the previous state-of-the-art results on all five datasets that we evaluate on. We then propose two new formulations of the person re-ID problem under resource-constraints, and show how our model can be used to effectively trade off accuracy and computation in the presence of resource constraints. Code and pre-trained models are available at https://github.com/mileyan/DARENet.

Utkarsh Mall, Cheng Perng Phoo, Meilin Kelsey Liu et al.

We introduce a method to train vision-language models for remote-sensing images without using any textual annotations. Our key insight is to use co-located internet imagery taken on the ground as an intermediary for connecting remote-sensing images and language. Specifically, we train an image encoder for remote sensing images to align with the image encoder of CLIP using a large amount of paired internet and satellite images. Our unsupervised approach enables the training of a first-of-its-kind large-scale vision language model (VLM) for remote sensing images at two different resolutions. We show that these VLMs enable zero-shot, open-vocabulary image classification, retrieval, segmentation and visual question answering for satellite images. On each of these tasks, our VLM trained without textual annotations outperforms existing VLMs trained with supervision, with gains of up to 20% for classification and 80% for segmentation.

Yihong Sun, Hao Zhou, Liangzhe Yuan et al. · deepmind

We explore a novel video creation experience, namely Video Creation by Demonstration. Given a demonstration video and a context image from a different scene, we generate a physically plausible video that continues naturally from the context image and carries out the action concepts from the demonstration. To enable this capability, we present $δ$-Diffusion, a self-supervised training approach that learns from unlabeled videos by conditional future frame prediction. Unlike most existing video generation controls that are based on explicit signals, we adopts the form of implicit latent control for maximal flexibility and expressiveness required by general videos. By leveraging a video foundation model with an appearance bottleneck design on top, we extract action latents from demonstration videos for conditioning the generation process with minimal appearance leakage. Empirically, $δ$-Diffusion outperforms related baselines in terms of both human preference and large-scale machine evaluations, and demonstrates potentials towards interactive world simulation. Sampled video generation results are available at https://delta-diffusion.github.io/.

Gene Chou, Charles Herrmann, Kyle Genova et al.

We address the problem of generating a 3D-consistent, navigable environment that is spatially grounded: a simulation of a real location. Existing video generative models can produce a plausible sequence that is consistent with a text (T2V) or image (I2V) prompt. However, the capability to reconstruct the real world under arbitrary weather conditions and dynamic object configurations is essential for downstream applications including autonomous driving and robotics simulation. To this end, we present CityRAG, a video generative model that leverages large corpora of geo-registered data as context to ground generation to the physical scene, while maintaining learned priors for complex motion and appearance changes. CityRAG relies on temporally unaligned training data, which teaches the model to semantically disentangle the underlying scene from its transient attributes. Our experiments demonstrate that CityRAG can generate coherent minutes-long, physically grounded video sequences, maintain weather and lighting conditions over thousands of frames, achieve loop closure, and navigate complex trajectories to reconstruct real-world geography.

Gene Chou, Kai Zhang, Sai Bi et al. · deepmind

We address the problem of generating videos from unposed internet photos. A handful of input images serve as keyframes, and our model interpolates between them to simulate a path moving between the cameras. Given random images, a model's ability to capture underlying geometry, recognize scene identity, and relate frames in terms of camera position and orientation reflects a fundamental understanding of 3D structure and scene layout. However, existing video models such as Luma Dream Machine fail at this task. We design a self-supervised method that takes advantage of the consistency of videos and variability of multiview internet photos to train a scalable, 3D-aware video model without any 3D annotations such as camera parameters. We validate that our method outperforms all baselines in terms of geometric and appearance consistency. We also show our model benefits applications that enable camera control, such as 3D Gaussian Splatting. Our results suggest that we can scale up scene-level 3D learning using only 2D data such as videos and multiview internet photos.

Utkarsh Mall, Cheng Perng Phoo, Mia Chiquier et al.

Visual data is used in numerous different scientific workflows ranging from remote sensing to ecology. As the amount of observation data increases, the challenge is not just to make accurate predictions but also to understand the underlying mechanisms for those predictions. Good interpretation is important in scientific workflows, as it allows for better decision-making by providing insights into the data. This paper introduces an automatic way of obtaining such interpretable-by-design models, by learning programs that interleave neural networks. We propose DiSciPLE (Discovering Scientific Programs using LLMs and Evolution) an evolutionary algorithm that leverages common sense and prior knowledge of large language models (LLMs) to create Python programs explaining visual data. Additionally, we propose two improvements: a program critic and a program simplifier to improve our method further to synthesize good programs. On three different real-world problems, DiSciPLE learns state-of-the-art programs on novel tasks with no prior literature. For example, we can learn programs with 35% lower error than the closest non-interpretable baseline for population density estimation.

Hangyu Zhou, Chia-Hsiang Kao, Cheng Perng Phoo et al.

Clouds in satellite imagery pose a significant challenge for downstream applications. A major challenge in current cloud removal research is the absence of a comprehensive benchmark and a sufficiently large and diverse training dataset. To address this problem, we introduce the largest public dataset -- $\textit{AllClear}$ for cloud removal, featuring 23,742 globally distributed regions of interest (ROIs) with diverse land-use patterns, comprising 4 million images in total. Each ROI includes complete temporal captures from the year 2022, with (1) multi-spectral optical imagery from Sentinel-2 and Landsat 8/9, (2) synthetic aperture radar (SAR) imagery from Sentinel-1, and (3) auxiliary remote sensing products such as cloud masks and land cover maps. We validate the effectiveness of our dataset by benchmarking performance, demonstrating the scaling law -- the PSNR rises from $28.47$ to $33.87$ with $30\times$ more data, and conducting ablation studies on the temporal length and the importance of individual modalities. This dataset aims to provide comprehensive coverage of the Earth's surface and promote better cloud removal results.