Serena Yeung

Orr Zohar, Kuan-Chieh Wang, Serena Yeung · stanford

Open World Object Detection (OWOD) is a new and challenging computer vision task that bridges the gap between classic object detection (OD) benchmarks and object detection in the real world. In addition to detecting and classifying seen/labeled objects, OWOD algorithms are expected to detect novel/unknown objects - which can be classified and incrementally learned. In standard OD, object proposals not overlapping with a labeled object are automatically classified as background. Therefore, simply applying OD methods to OWOD fails as unknown objects would be predicted as background. The challenge of detecting unknown objects stems from the lack of supervision in distinguishing unknown objects and background object proposals. Previous OWOD methods have attempted to overcome this issue by generating supervision using pseudo-labeling - however, unknown object detection has remained low. Probabilistic/generative models may provide a solution for this challenge. Herein, we introduce a novel probabilistic framework for objectness estimation, where we alternate between probability distribution estimation and objectness likelihood maximization of known objects in the embedded feature space - ultimately allowing us to estimate the objectness probability of different proposals. The resulting Probabilistic Objectness transformer-based open-world detector, PROB, integrates our framework into traditional object detection models, adapting them for the open-world setting. Comprehensive experiments on OWOD benchmarks show that PROB outperforms all existing OWOD methods in both unknown object detection ($\sim 2\times$ unknown recall) and known object detection ($\sim 10\%$ mAP). Our code will be made available upon publication at https://github.com/orrzohar/PROB.

Weixin Liang, Yuhui Zhang, Yongchan Kwon et al. · stanford

We present modality gap, an intriguing geometric phenomenon of the representation space of multi-modal models. Specifically, we show that different data modalities (e.g. images and text) are embedded at arm's length in their shared representation in multi-modal models such as CLIP. Our systematic analysis demonstrates that this gap is caused by a combination of model initialization and contrastive learning optimization. In model initialization, we show empirically and theoretically that the representation of a common deep neural network is restricted to a narrow cone. As a consequence, in a multi-modal model with two encoders, the representations of the two modalities are clearly apart when the model is initialized. During optimization, contrastive learning keeps the different modalities separate by a certain distance, which is influenced by the temperature parameter in the loss function. Our experiments further demonstrate that varying the modality gap distance has a significant impact in improving the model's downstream zero-shot classification performance and fairness. Our code and data are available at https://modalitygap.readthedocs.io/

Mark Mazumder, Colby Banbury, Xiaozhe Yao et al.

Machine learning research has long focused on models rather than datasets, and prominent datasets are used for common ML tasks without regard to the breadth, difficulty, and faithfulness of the underlying problems. Neglecting the fundamental importance of data has given rise to inaccuracy, bias, and fragility in real-world applications, and research is hindered by saturation across existing dataset benchmarks. In response, we present DataPerf, a community-led benchmark suite for evaluating ML datasets and data-centric algorithms. We aim to foster innovation in data-centric AI through competition, comparability, and reproducibility. We enable the ML community to iterate on datasets, instead of just architectures, and we provide an open, online platform with multiple rounds of challenges to support this iterative development. The first iteration of DataPerf contains five benchmarks covering a wide spectrum of data-centric techniques, tasks, and modalities in vision, speech, acquisition, debugging, and diffusion prompting, and we support hosting new contributed benchmarks from the community. The benchmarks, online evaluation platform, and baseline implementations are open source, and the MLCommons Association will maintain DataPerf to ensure long-term benefits to academia and industry.

Yuhui Zhang, Jeff Z. HaoChen, Shih-Cheng Huang et al. · stanford

Recent multi-modal contrastive learning models have demonstrated the ability to learn an embedding space suitable for building strong vision classifiers, by leveraging the rich information in large-scale image-caption datasets. Our work highlights a distinct advantage of this multi-modal embedding space: the ability to diagnose vision classifiers through natural language. The traditional process of diagnosing model behaviors in deployment settings involves labor-intensive data acquisition and annotation. Our proposed method can discover high-error data slices, identify influential attributes and further rectify undesirable model behaviors, without requiring any visual data. Through a combination of theoretical explanation and empirical verification, we present conditions under which classifiers trained on embeddings from one modality can be equivalently applied to embeddings from another modality. On a range of image datasets with known error slices, we demonstrate that our method can effectively identify the error slices and influential attributes, and can further use language to rectify failure modes of the classifier.

Yuhui Zhang, Shih-Cheng Huang, Zhengping Zhou et al. · stanford

Given the prevalence of 3D medical imaging technologies such as MRI and CT that are widely used in diagnosing and treating diverse diseases, 3D segmentation is one of the fundamental tasks of medical image analysis. Recently, Transformer-based models have started to achieve state-of-the-art performances across many vision tasks, through pre-training on large-scale natural image benchmark datasets. While works on medical image analysis have also begun to explore Transformer-based models, there is currently no optimal strategy to effectively leverage pre-trained Transformers, primarily due to the difference in dimensionality between 2D natural images and 3D medical images. Existing solutions either split 3D images into 2D slices and predict each slice independently, thereby losing crucial depth-wise information, or modify the Transformer architecture to support 3D inputs without leveraging pre-trained weights. In this work, we use a simple yet effective weight inflation strategy to adapt pre-trained Transformers from 2D to 3D, retaining the benefit of both transfer learning and depth information. We further investigate the effectiveness of transfer from different pre-training sources and objectives. Our approach achieves state-of-the-art performances across a broad range of 3D medical image datasets, and can become a standard strategy easily utilized by all work on Transformer-based models for 3D medical images, to maximize performance.

Zhenzhen Weng, Kuan-Chieh Wang, Angjoo Kanazawa et al. · stanford

The ability to perceive 3D human bodies from a single image has a multitude of applications ranging from entertainment and robotics to neuroscience and healthcare. A fundamental challenge in human mesh recovery is in collecting the ground truth 3D mesh targets required for training, which requires burdensome motion capturing systems and is often limited to indoor laboratories. As a result, while progress is made on benchmark datasets collected in these restrictive settings, models fail to generalize to real-world "in-the-wild" scenarios due to distribution shifts. We propose Domain Adaptive 3D Pose Augmentation (DAPA), a data augmentation method that enhances the model's generalization ability in in-the-wild scenarios. DAPA combines the strength of methods based on synthetic datasets by getting direct supervision from the synthesized meshes, and domain adaptation methods by using ground truth 2D keypoints from the target dataset. We show quantitatively that finetuning with DAPA effectively improves results on benchmarks 3DPW and AGORA. We further demonstrate the utility of DAPA on a challenging dataset curated from videos of real-world parent-child interaction.

Orr Zohar, Shih-Cheng Huang, Kuan-Chieh Wang et al.

Pre-trained multi-modal vision-language models (VLMs) are becoming increasingly popular due to their exceptional performance on downstream vision applications, particularly in the few- and zero-shot settings. However, selecting the best-performing VLM for some downstream applications is non-trivial, as it is dataset and task-dependent. Meanwhile, the exhaustive evaluation of all available VLMs on a novel application is not only time and computationally demanding but also necessitates the collection of a labeled dataset for evaluation. As the number of open-source VLM variants increases, there is a need for an efficient model selection strategy that does not require access to a curated evaluation dataset. This paper proposes a novel task and benchmark for efficiently evaluating VLMs' zero-shot performance on downstream applications without access to the downstream task dataset. Specifically, we introduce a new task LOVM: Language-Only Vision Model Selection, where methods are expected to perform both model selection and performance prediction based solely on a text description of the desired downstream application. We then introduced an extensive LOVM benchmark consisting of ground-truth evaluations of 35 pre-trained VLMs and 23 datasets, where methods are expected to rank the pre-trained VLMs and predict their zero-shot performance.

Alexander Ke, Shih-Cheng Huang, Chloe P O'Connell et al.

Pretraining on large natural image classification datasets such as ImageNet has aided model development on data-scarce 2D medical tasks. 3D medical tasks often have much less data than 2D medical tasks, prompting practitioners to rely on pretrained 2D models to featurize slices. However, these 2D models have been surpassed by 3D models on 3D computer vision benchmarks since they do not natively leverage cross-sectional or temporal information. In this study, we explore whether natural video pretraining for 3D models can enable higher performance on smaller datasets for 3D medical tasks. We demonstrate video pretraining improves the average performance of seven 3D models on two chest CT datasets, regardless of finetuning dataset size, and that video pretraining allows 3D models to outperform 2D baselines. Lastly, we observe that pretraining on the large-scale out-of-domain Kinetics dataset improves performance more than pretraining on a typically-sized in-domain CT dataset. Our results show consistent benefits of video pretraining across a wide array of architectures, tasks, and training dataset sizes, supporting a shift from small-scale in-domain pretraining to large-scale out-of-domain pretraining for 3D medical tasks. Our code is available at: https://github.com/rajpurkarlab/chest-ct-pretraining

Kuan-Chieh Wang, Zhenzhen Weng, Maria Xenochristou et al. · stanford

The task of reconstructing 3D human motion has wideranging applications. The gold standard Motion capture (MoCap) systems are accurate but inaccessible to the general public due to their cost, hardware and space constraints. In contrast, monocular human mesh recovery (HMR) methods are much more accessible than MoCap as they take single-view videos as inputs. Replacing the multi-view Mo- Cap systems with a monocular HMR method would break the current barriers to collecting accurate 3D motion thus making exciting applications like motion analysis and motiondriven animation accessible to the general public. However, performance of existing HMR methods degrade when the video contains challenging and dynamic motion that is not in existing MoCap datasets used for training. This reduces its appeal as dynamic motion is frequently the target in 3D motion recovery in the aforementioned applications. Our study aims to bridge the gap between monocular HMR and multi-view MoCap systems by leveraging information shared across multiple video instances of the same action. We introduce the Neural Motion (NeMo) field. It is optimized to represent the underlying 3D motions across a set of videos of the same action. Empirically, we show that NeMo can recover 3D motion in sports using videos from the Penn Action dataset, where NeMo outperforms existing HMR methods in terms of 2D keypoint detection. To further validate NeMo using 3D metrics, we collected a small MoCap dataset mimicking actions in Penn Action,and show that NeMo achieves better 3D reconstruction compared to various baselines.

James Burgess, Kuan-Chieh Wang, Serena Yeung-Levy

Text-to-image diffusion models generate impressive and realistic images, but do they learn to represent the 3D world from only 2D supervision? We demonstrate that yes, certain 3D scene representations are encoded in the text embedding space of models like Stable Diffusion. Our approach, Viewpoint Neural Textual Inversion (ViewNeTI), is to discover 3D view tokens; these tokens control the 3D viewpoint - the rendering pose in a scene - of generated images. Specifically, we train a small neural mapper to take continuous camera viewpoint parameters and predict a view token (a word embedding). This token conditions diffusion generation via cross-attention to produce images with the desired camera viewpoint. Using ViewNeTI as an evaluation tool, we report two findings: first, the text latent space has a continuous view-control manifold for particular 3D scenes; second, we find evidence for a generalized view-control manifold for all scenes. We conclude that since the view token controls the 3D `rendering' viewpoint, there is likely a scene representation embedded in frozen 2D diffusion models. Finally, we exploit the 3D scene representations for 3D vision tasks, namely, view-controlled text-to-image generation, and novel view synthesis from a single image, where our approach sets state-of-the-art for LPIPS. Code available at https://github.com/jmhb0/view_neti

Ali Mottaghi, Aidean Sharghi, Serena Yeung et al.

Automatic surgical activity recognition enables more intelligent surgical devices and a more efficient workflow. Integration of such technology in new operating rooms has the potential to improve care delivery to patients and decrease costs. Recent works have achieved a promising performance on surgical activity recognition; however, the lack of generalizability of these models is one of the critical barriers to the wide-scale adoption of this technology. In this work, we study the generalizability of surgical activity recognition models across operating rooms. We propose a new domain adaptation method to improve the performance of the surgical activity recognition model in a new operating room for which we only have unlabeled videos. Our approach generates pseudo labels for unlabeled video clips that it is confident about and trains the model on the augmented version of the clips. We extend our method to a semi-supervised domain adaptation setting where a small portion of the target domain is also labeled. In our experiments, our proposed method consistently outperforms the baselines on a dataset of more than 480 long surgical videos collected from two operating rooms.

Shih-Cheng Huang, Zepeng Huo, Ethan Steinberg et al.

Synthesizing information from multiple data sources plays a crucial role in the practice of modern medicine. Current applications of artificial intelligence in medicine often focus on single-modality data due to a lack of publicly available, multimodal medical datasets. To address this limitation, we introduce INSPECT, which contains de-identified longitudinal records from a large cohort of patients at risk for pulmonary embolism (PE), along with ground truth labels for multiple outcomes. INSPECT contains data from 19,402 patients, including CT images, radiology report impression sections, and structured electronic health record (EHR) data (i.e. demographics, diagnoses, procedures, vitals, and medications). Using INSPECT, we develop and release a benchmark for evaluating several baseline modeling approaches on a variety of important PE related tasks. We evaluate image-only, EHR-only, and multimodal fusion models. Trained models and the de-identified dataset are made available for non-commercial use under a data use agreement. To the best of our knowledge, INSPECT is the largest multimodal dataset integrating 3D medical imaging and EHR for reproducible methods evaluation and research.

Zhenzhen Weng, Laura Bravo-Sánchez, Serena Yeung-Levy

Recent text-to-image generative models have exhibited remarkable abilities in generating high-fidelity and photo-realistic images. However, despite the visually impressive results, these models often struggle to preserve plausible human structure in the generations. Due to this reason, while generative models have shown promising results in aiding downstream image recognition tasks by generating large volumes of synthetic data, they are not suitable for improving downstream human pose perception and understanding. In this work, we propose a Diffusion model with Human Pose Correction (Diffusion-HPC), a text-conditioned method that generates photo-realistic images with plausible posed humans by injecting prior knowledge about human body structure. Our generated images are accompanied by 3D meshes that serve as ground truths for improving Human Mesh Recovery tasks, where a shortage of 3D training data has long been an issue. Furthermore, we show that Diffusion-HPC effectively improves the realism of human generations under varying conditioning strategies.

Jeffrey Gu, Kuan-Chieh Wang, Serena Yeung

Neural fields, which represent signals as a function parameterized by a neural network, are a promising alternative to traditional discrete vector or grid-based representations. Compared to discrete representations, neural representations both scale well with increasing resolution, are continuous, and can be many-times differentiable. However, given a dataset of signals that we would like to represent, having to optimize a separate neural field for each signal is inefficient, and cannot capitalize on shared information or structures among signals. Existing generalization methods view this as a meta-learning problem and employ gradient-based meta-learning to learn an initialization which is then fine-tuned with test-time optimization, or learn hypernetworks to produce the weights of a neural field. We instead propose a new paradigm that views the large-scale training of neural representations as a part of a partially-observed neural process framework, and leverage neural process algorithms to solve this task. We demonstrate that this approach outperforms both state-of-the-art gradient-based meta-learning approaches and hypernetwork approaches.

Juan Manuel Zambrano Chaves, Shih-Cheng Huang, Yanbo Xu et al. · microsoft-research

The scaling laws and extraordinary performance of large foundation models motivate the development and utilization of such models in biomedicine. However, despite early promising results on some biomedical benchmarks, there are still major challenges that need to be addressed before these models can be used in real-world clinics. Frontier general-domain models such as GPT-4V still have significant performance gaps in multimodal biomedical applications. More importantly, less-acknowledged pragmatic issues, including accessibility, model cost, and tedious manual evaluation make it hard for clinicians to use state-of-the-art large models directly on private patient data. Here, we explore training open-source small multimodal models (SMMs) to bridge competency gaps for unmet clinical needs in radiology. To maximize data efficiency, we adopt a modular approach by incorporating state-of-the-art pre-trained models for image and text modalities, and focusing on training a lightweight adapter to ground each modality to the text embedding space, as exemplified by LLaVA-Med. For training, we assemble a large dataset of over 697 thousand radiology image-text pairs. For evaluation, we propose CheXprompt, a GPT-4-based metric for factuality evaluation, and demonstrate its parity with expert evaluation. For best practice, we conduct a systematic ablation study on various choices in data engineering and multimodal training. The resulting LlaVA-Rad (7B) model attains state-of-the-art results on standard radiology tasks such as report generation and cross-modal retrieval, even outperforming much larger models such as GPT-4V and Med-PaLM M (84B). The inference of LlaVA-Rad is fast and can be performed on a single V100 GPU in private settings, offering a promising state-of-the-art tool for real-world clinical applications.

James Burgess, Xiaohan Wang, Yuhui Zhang et al. · stanford

How do two individuals differ when performing the same action? In this work, we introduce Video Action Differencing (VidDiff), the novel task of identifying subtle differences between videos of the same action, which has many applications, such as coaching and skill learning. To enable development on this new task, we first create VidDiffBench, a benchmark dataset containing 549 video pairs, with human annotations of 4,469 fine-grained action differences and 2,075 localization timestamps indicating where these differences occur. Our experiments demonstrate that VidDiffBench poses a significant challenge for state-of-the-art large multimodal models (LMMs), such as GPT-4o and Qwen2-VL. By analyzing failure cases of LMMs on VidDiffBench, we highlight two key challenges for this task: localizing relevant sub-actions over two videos and fine-grained frame comparison. To overcome these, we propose the VidDiff method, an agentic workflow that breaks the task into three stages: action difference proposal, keyframe localization, and frame differencing, each stage utilizing specialized foundation models. To encourage future research in this new task, we release the benchmark at https://huggingface.co/datasets/jmhb/VidDiffBench and code at http://jmhb0.github.io/viddiff.

Josiah Aklilu, Xiaohan Wang, Serena Yeung-Levy

Precise action localization in untrimmed video is vital for fields such as professional sports and minimally invasive surgery, where the delineation of particular motions in recordings can dramatically enhance analysis. But in many cases, large scale datasets with video-label pairs for localization are unavailable, limiting the opportunity to fine-tune video-understanding models. Recent developments in large vision-language models (LVLM) address this need with impressive zero-shot capabilities in a variety of video understanding tasks. However, the adaptation of LVLMs, with their powerful visual question answering capabilities, to zero-shot localization in long-form video is still relatively unexplored. To this end, we introduce a true Zero-shot Action Localization method (ZEAL). Specifically, we leverage the built-in action knowledge of a large language model (LLM) to inflate actions into detailed descriptions of the archetypal start and end of the action. These descriptions serve as queries to LVLM for generating frame-level confidence scores which can be aggregated to produce localization outputs. The simplicity and flexibility of our method lends it amenable to more capable LVLMs as they are developed, and we demonstrate remarkable results in zero-shot action localization on a challenging benchmark, without any training. Our code is publicly available at $\href{https://github.com/josaklil-ai/zeal}{github.com/josaklil-ai/zeal}$.

Andong Deng, Taojiannan Yang, Shoubin Yu et al.

Large Multimodal Models (LMMs) have achieved remarkable progress across various capabilities; however, complex video reasoning in the scientific domain remains a significant and challenging frontier. Current video benchmarks predominantly target general scenarios where perception/recognition is heavily relied on, while with relatively simple reasoning tasks, leading to saturation and thus failing to effectively evaluate advanced multimodal cognitive skills. To address this critical gap, we introduce SciVideoBench, a rigorous benchmark specifically designed to assess advanced video reasoning in scientific contexts. SciVideoBench consists of 1,000 carefully crafted multiple-choice questions derived from cutting-edge scientific experimental videos spanning over 25 specialized academic subjects and verified by a semi-automatic system. Each question demands sophisticated domain-specific knowledge, precise spatiotemporal perception, and intricate logical reasoning, effectively challenging models' higher-order cognitive abilities. Our evaluation highlights significant performance deficits in state-of-the-art proprietary and open-source LMMs, including Gemini 2.5 Pro and Qwen2.5-VL, indicating substantial room for advancement in video reasoning capabilities. Detailed analyses of critical factors such as reasoning complexity and visual grounding provide valuable insights and clear direction for future developments in LMMs, driving the evolution of truly capable multimodal AI co-scientists. We hope SciVideoBench could fit the interests of the community and help to push the boundary of cutting-edge AI for border science.

Sanket Rajan Gupte, Josiah Aklilu, Jeffrey J. Nirschl et al.

Foundation vision or vision-language models are trained on large unlabeled or noisy data and learn robust representations that can achieve impressive zero- or few-shot performance on diverse tasks. Given these properties, they are a natural fit for active learning (AL), which aims to maximize labeling efficiency. However, the full potential of foundation models has not been explored in the context of AL, specifically in the low-budget regime. In this work, we evaluate how foundation models influence three critical components of effective AL, namely, 1) initial labeled pool selection, 2) ensuring diverse sampling, and 3) the trade-off between representative and uncertainty sampling. We systematically study how the robust representations of foundation models (DINOv2, OpenCLIP) challenge existing findings in active learning. Our observations inform the principled construction of a new simple and elegant AL strategy that balances uncertainty estimated via dropout with sample diversity. We extensively test our strategy on many challenging image classification benchmarks, including natural images as well as out-of-domain biomedical images that are relatively understudied in the AL literature. We also provide a highly performant and efficient implementation of modern AL strategies (including our method) at https://github.com/sanketx/AL-foundation-models.

Orr Zohar, Alejandro Lozano, Shelly Goel et al.

Object detection is integral to a bevy of real-world applications, from robotics to medical image analysis. To be used reliably in such applications, models must be capable of handling unexpected - or novel - objects. The open world object detection (OWD) paradigm addresses this challenge by enabling models to detect unknown objects and learn discovered ones incrementally. However, OWD method development is hindered due to the stringent benchmark and task definitions. These definitions effectively prohibit foundation models. Here, we aim to relax these definitions and investigate the utilization of pre-trained foundation models in OWD. First, we show that existing benchmarks are insufficient in evaluating methods that utilize foundation models, as even naive integration methods nearly saturate these benchmarks. This result motivated us to curate a new and challenging benchmark for these models. Therefore, we introduce a new benchmark that includes five real-world application-driven datasets, including challenging domains such as aerial and surgical images, and establish baselines. We exploit the inherent connection between classes in application-driven datasets and introduce a novel method, Foundation Object detection Model for the Open world, or FOMO, which identifies unknown objects based on their shared attributes with the base known objects. FOMO has ~3x unknown object mAP compared to baselines on our benchmark. However, our results indicate a significant place for improvement - suggesting a great research opportunity in further scaling object detection methods to real-world domains. Our code and benchmark are available at https://orrzohar.github.io/projects/fomo/.

Devanish N. Kamtam, Joseph B. Shrager, Satya Deepya Malla et al.

Background: We evaluate SAM 2 for surgical scene understanding by examining its semantic segmentation capabilities for organs/tissues both in zero-shot scenarios and after fine-tuning. Methods: We utilized five public datasets to evaluate and fine-tune SAM 2 for segmenting anatomical tissues in surgical videos/images. Fine-tuning was applied to the image encoder and mask decoder. We limited training subsets from 50 to 400 samples per class to better model real-world constraints with data acquisition. The impact of dataset size on fine-tuning performance was evaluated with weighted mean Dice coefficient (WMDC), and the results were also compared against previously reported state-of-the-art (SOTA) results. Results: SurgiSAM 2, a fine-tuned SAM 2 model, demonstrated significant improvements in segmentation performance, achieving a 17.9% relative WMDC gain compared to the baseline SAM 2. Increasing prompt points from 1 to 10 and training data scale from 50/class to 400/class enhanced performance; the best WMDC of 0.92 on the validation subset was achieved with 10 prompt points and 400 samples per class. On the test subset, this model outperformed prior SOTA methods in 24/30 (80%) of the classes with a WMDC of 0.91 using 10-point prompts. Notably, SurgiSAM 2 generalized effectively to unseen organ classes, achieving SOTA on 7/9 (77.8%) of them. Conclusion: SAM 2 achieves remarkable zero-shot and fine-tuned performance for surgical scene segmentation, surpassing prior SOTA models across several organ classes of diverse datasets. This suggests immense potential for enabling automated/semi-automated annotation pipelines, thereby decreasing the burden of annotations facilitating several surgical applications.

Zeyu Wang, Zhenzhen Weng, Serena Yeung-Levy

Conventional approaches to human mesh recovery predominantly employ a region-based strategy. This involves initially cropping out a human-centered region as a preprocessing step, with subsequent modeling focused on this zoomed-in image. While effective for single figures, this pipeline poses challenges when dealing with images featuring multiple individuals, as different people are processed separately, often leading to inaccuracies in relative positioning. Despite the advantages of adopting a whole-image-based approach to address this limitation, early efforts in this direction have fallen short in performance compared to recent region-based methods. In this work, we advocate for this under-explored area of modeling all people at once, emphasizing its potential for improved accuracy in multi-person scenarios through considering all individuals simultaneously and leveraging the overall context and interactions. We introduce a new model with a streamlined transformer-based design, featuring three critical design choices: multi-scale feature incorporation, focused attention mechanisms, and relative joint supervision. Our proposed model demonstrates a significant performance improvement, surpassing state-of-the-art region-based and whole-image-based methods on various benchmarks involving multiple individuals.

Ali Mottaghi, Mohammad Abdullah Jamal, Serena Yeung et al.

Semi-supervised domain adaptation (SSDA) presents a critical hurdle in computer vision, especially given the frequent scarcity of labeled data in real-world settings. This scarcity often causes foundation models, trained on extensive datasets, to underperform when applied to new domains. AdaEmbed, our newly proposed methodology for SSDA, offers a promising solution to these challenges. Leveraging the potential of unlabeled data, AdaEmbed facilitates the transfer of knowledge from a labeled source domain to an unlabeled target domain by learning a shared embedding space. By generating accurate and uniform pseudo-labels based on the established embedding space, the model overcomes the limitations of conventional SSDA, thus enhancing performance significantly. Our method's effectiveness is validated through extensive experiments on benchmark datasets such as DomainNet, Office-Home, and VisDA-C, where AdaEmbed consistently outperforms all the baselines, setting a new state of the art for SSDA. With its straightforward implementation and high data efficiency, AdaEmbed stands out as a robust and pragmatic solution for real-world scenarios, where labeled data is scarce. To foster further research and application in this area, we are sharing the codebase of our unified framework for semi-supervised domain adaptation.

Mark Endo, Serena Yeung-Levy

Scaling up multimodal models has enabled remarkable advances in visual understanding and reasoning, but practical demands call for smaller, efficient systems. In this work, we conduct a principled analysis of downscaling intelligence in multimodal models, examining how reduced large language model (LLM) capacity affects multimodal capabilities. Our initial findings reveal an interesting trend: LLM downscaling disproportionately affects visual capabilities, rather than abilities inherited from the LLM. We then examine whether this drop mainly reflects the expected decline in visual reasoning or a more fundamental loss of perceptual abilities. Isolating the effect of LLM downscaling on perception, we find performance still drops sharply, often matching or exceeding the impact on reasoning. To address this bottleneck, we introduce visual extraction tuning, which explicitly trains the model to extract instruction-relevant visual details consistently across tasks. With these extracted visual details, we then apply step-by-step reasoning to generate answers. Together, these components form our Extract+Think approach, setting a new standard for efficiency and performance in this space.

Liangyu Chen, James Burgess, Jeffrey J Nirschl et al.

Imaging technologies are fundamental to biomedical research and modern medicine, requiring analysis of high-resolution images across various modalities. While multimodal large language models (MLLMs) show promise for biomedical image analysis, most are designed for low-resolution images from general-purpose datasets, risking critical information loss. We investigate how image resolution affects MLLM performance in biomedical applications and demonstrate that: (1) native-resolution training and inference significantly improve performance across multiple tasks, (2) misalignment between training and inference resolutions severely degrades performance, and (3) mixed-resolution training effectively mitigates misalignment and balances computational constraints with performance requirements. Based on these findings, we recommend prioritizing native-resolution inference and mixed-resolution datasets to optimize biomedical MLLMs for transformative impact in scientific research and clinical applications.

Yiming Liu, Yuhui Zhang, Dhruba Ghosh et al. · stanford

CLIP outperforms self-supervised models like DINO as vision encoders for vision-language models (VLMs), but it remains unclear whether this advantage stems from CLIP's language supervision or its much larger training data. To disentangle these factors, we pre-train CLIP and DINO under controlled settings -- using the same architecture, dataset, and training configuration -- achieving similar ImageNet accuracy. Embedding analysis shows that CLIP captures high-level semantics (e.g., object categories, text), while DINO is more responsive to low-level features like colors and styles. When integrated into VLMs and evaluated on 20 VQA benchmarks, CLIP excels at text-intensive tasks, while DINO slightly outperforms on vision-centric ones. Variants of language supervision (e.g., sigmoid loss, pre-trained language encoders) yield limited gains. Our findings provide scientific insights into vision encoder design and its impact on VLM performance.

Anita Rau, Josiah Aklilu, F. Christopher Holsinger et al.

Neural Radiance Fields (NeRFs) are trained to minimize the rendering loss of predicted viewpoints. However, the photometric loss often does not provide enough information to disambiguate between different possible geometries yielding the same image. Previous work has thus incorporated depth supervision during NeRF training, leveraging dense predictions from pre-trained depth networks as pseudo-ground truth. While these depth priors are assumed to be perfect once filtered for noise, in practice, their accuracy is more challenging to capture. This work proposes a novel approach to uncertainty in depth priors for NeRF supervision. Instead of using custom-trained depth or uncertainty priors, we use off-the-shelf pretrained diffusion models to predict depth and capture uncertainty during the denoising process. Because we know that depth priors are prone to errors, we propose to supervise the ray termination distance distribution with Earth Mover's Distance instead of enforcing the rendered depth to replicate the depth prior exactly through L2-loss. Our depth-guided NeRF outperforms all baselines on standard depth metrics by a large margin while maintaining performance on photometric measures.

Elaine Sui, Xiaohan Wang, Serena Yeung-Levy

Advancements in vision-language models (VLMs) have propelled the field of computer vision, particularly in the zero-shot learning setting. Despite their promise, the effectiveness of these models often diminishes due to domain shifts in test environments. To address this, we introduce the Test-Time Prototype Shifting (TPS) framework, a pioneering approach designed to adapt VLMs to test datasets using unlabeled test inputs. Our method is based on the notion of modulating per-class prototypes in the shared embedding space. By pre-computing and caching prototypes generated with the pre-trained text encoder, TPS not only facilitates optimization-free prototype reuse for subsequent predictions but also enables seamless integration with current advancements in prompt engineering. At test-time, TPS dynamically learns shift vectors for each prototype based solely on the given test sample, effectively bridging the domain gap and enhancing classification accuracy. A notable aspect of our framework is its significantly reduced memory and computational demands when compared to conventional text-prompt tuning methods. Extensive evaluations across 15 image classification datasets involving natural distribution shifts and cross-dataset generalization, as well as in context-dependent visual reasoning, demonstrate TPS's superior performance, achieving state-of-the-art results while reducing resource requirements.

Yuhui Zhang, Michihiro Yasunaga, Zhengping Zhou et al.

Language models have been shown to exhibit positive scaling, where performance improves as models are scaled up in terms of size, compute, or data. In this work, we introduce NeQA, a dataset consisting of questions with negation in which language models do not exhibit straightforward positive scaling. We show that this task can exhibit inverse scaling, U-shaped scaling, or positive scaling, and the three scaling trends shift in this order as we use more powerful prompting methods or model families. We hypothesize that solving NeQA depends on two subtasks: question answering (task 1) and negation understanding (task 2). We find that task 1 has linear scaling, while task 2 has sigmoid-shaped scaling with an emergent transition point, and composing these two scaling trends yields the final scaling trend of NeQA. Our work reveals and provides a way to analyze the complex scaling trends of language models.

Zhenzhen Weng, Zeyu Wang, Serena Yeung

Recent advancements in text-to-image generation have enabled significant progress in zero-shot 3D shape generation. This is achieved by score distillation, a methodology that uses pre-trained text-to-image diffusion models to optimize the parameters of a 3D neural presentation, e.g. Neural Radiance Field (NeRF). While showing promising results, existing methods are often not able to preserve the geometry of complex shapes, such as human bodies. To address this challenge, we present ZeroAvatar, a method that introduces the explicit 3D human body prior to the optimization process. Specifically, we first estimate and refine the parameters of a parametric human body from a single image. Then during optimization, we use the posed parametric body as additional geometry constraint to regularize the diffusion model as well as the underlying density field. Lastly, we propose a UV-guided texture regularization term to further guide the completion of texture on invisible body parts. We show that ZeroAvatar significantly enhances the robustness and 3D consistency of optimization-based image-to-3D avatar generation, outperforming existing zero-shot image-to-3D methods.

Pascal Mettes, Mina Ghadimi Atigh, Martin Keller-Ressel et al.

Deep representation learning is a ubiquitous part of modern computer vision. While Euclidean space has been the de facto standard manifold for learning visual representations, hyperbolic space has recently gained rapid traction for learning in computer vision. Specifically, hyperbolic learning has shown a strong potential to embed hierarchical structures, learn from limited samples, quantify uncertainty, add robustness, limit error severity, and more. In this paper, we provide a categorization and in-depth overview of current literature on hyperbolic learning for computer vision. We research both supervised and unsupervised literature and identify three main research themes in each direction. We outline how hyperbolic learning is performed in all themes and discuss the main research problems that benefit from current advances in hyperbolic learning for computer vision. Moreover, we provide a high-level intuition behind hyperbolic geometry and outline open research questions to further advance research in this direction.

Emmett D. Goodman, Krishna K. Patel, Yilun Zhang et al.

Open procedures represent the dominant form of surgery worldwide. Artificial intelligence (AI) has the potential to optimize surgical practice and improve patient outcomes, but efforts have focused primarily on minimally invasive techniques. Our work overcomes existing data limitations for training AI models by curating, from YouTube, the largest dataset of open surgical videos to date: 1997 videos from 23 surgical procedures uploaded from 50 countries. Using this dataset, we developed a multi-task AI model capable of real-time understanding of surgical behaviors, hands, and tools - the building blocks of procedural flow and surgeon skill. We show that our model generalizes across diverse surgery types and environments. Illustrating this generalizability, we directly applied our YouTube-trained model to analyze open surgeries prospectively collected at an academic medical center and identified kinematic descriptors of surgical skill related to efficiency of hand motion. Our Annotated Videos of Open Surgery (AVOS) dataset and trained model will be made available for further development of surgical AI.

Julia Gong, F. Christopher Holsinger, Serena Yeung

We consider the task of semi-supervised video object segmentation (VOS). Our approach mitigates shortcomings in previous VOS work by addressing detail preservation and temporal consistency using visual warping. In contrast to prior work that uses full optical flow, we introduce a new foreground-targeted visual warping approach that learns flow fields from VOS data. We train a flow module to capture detailed motion between frames using two weakly-supervised losses. Our object-focused approach of warping previous foreground object masks to their positions in the target frame enables detailed mask refinement with fast runtimes without using extra flow supervision. It can also be integrated directly into state-of-the-art segmentation networks. On the DAVIS17 and YouTubeVOS benchmarks, we outperform state-of-the-art offline methods that do not use extra data, as well as many online methods that use extra data. Qualitatively, we also show our approach produces segmentations with high detail and temporal consistency.

Jeffrey Gu, Serena Yeung

Creating representations of shapes that are invari-ant to isometric or almost-isometric transforma-tions has long been an area of interest in shape anal-ysis, since enforcing invariance allows the learningof more effective and robust shape representations.Most existing invariant shape representations arehandcrafted, and previous work on learning shaperepresentations do not focus on producing invariantrepresentations. To solve the problem of learningunsupervised invariant shape representations, weuse contrastive learning, which produces discrimi-native representations through learning invarianceto user-specified data augmentations. To producerepresentations that are specifically isometry andalmost-isometry invariant, we propose new dataaugmentations that randomly sample these transfor-mations. We show experimentally that our methodoutperforms previous unsupervised learning ap-proaches in both effectiveness and robustness.

Joy Hsu, Wah Chiu, Serena Yeung

In the biomedical domain, there is an abundance of dense, complex data where objects of interest may be challenging to detect or constrained by limits of human knowledge. Labelled domain specific datasets for supervised tasks are often expensive to obtain, and furthermore discovery of novel distinct objects may be desirable for unbiased scientific discovery. Therefore, we propose leveraging the wealth of annotations in benchmark computer vision datasets to conduct unsupervised instance segmentation for diverse biomedical datasets. The key obstacle is thus overcoming the large domain shift from common to biomedical images. We propose a Domain Adaptive Region-based Convolutional Neural Network (DARCNN), that adapts knowledge of object definition from COCO, a large labelled vision dataset, to multiple biomedical datasets. We introduce a domain separation module, a self-supervised representation consistency loss, and an augmented pseudo-labelling stage within DARCNN to effectively perform domain adaptation across such large domain shifts. We showcase DARCNN's performance for unsupervised instance segmentation on numerous biomedical datasets.

Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik et al.

Instance segmentation is an active topic in computer vision that is usually solved by using supervised learning approaches over very large datasets composed of object level masks. Obtaining such a dataset for any new domain can be very expensive and time-consuming. In addition, models trained on certain annotated categories do not generalize well to unseen objects. The goal of this paper is to propose a method that can perform unsupervised discovery of long-tail categories in instance segmentation, through learning instance embeddings of masked regions. Leveraging rich relationship and hierarchical structure between objects in the images, we propose self-supervised losses for learning mask embeddings. Trained on COCO dataset without additional annotations of the long-tail objects, our model is able to discover novel and more fine-grained objects than the common categories in COCO. We show that the model achieves competitive quantitative results on LVIS as compared to the supervised and partially supervised methods.

Michael Zhang, Karan Sapra, Sanja Fidler et al.

While federated learning traditionally aims to train a single global model across decentralized local datasets, one model may not always be ideal for all participating clients. Here we propose an alternative, where each client only federates with other relevant clients to obtain a stronger model per client-specific objectives. To achieve this personalization, rather than computing a single model average with constant weights for the entire federation as in traditional FL, we efficiently calculate optimal weighted model combinations for each client, based on figuring out how much a client can benefit from another's model. We do not assume knowledge of any underlying data distributions or client similarities, and allow each client to optimize for arbitrary target distributions of interest, enabling greater flexibility for personalization. We evaluate and characterize our method on a variety of federated settings, datasets, and degrees of local data heterogeneity. Our method outperforms existing alternatives, while also enabling new features for personalized FL such as transfer outside of local data distributions.

Michael Zhang, Xiaotian Cheng, Daniel Copeland et al.

Open, or non-laparoscopic surgery, represents the vast majority of all operating room procedures, but few tools exist to objectively evaluate these techniques at scale. Current efforts involve human expert-based visual assessment. We leverage advances in computer vision to introduce an automated approach to video analysis of surgical execution. A state-of-the-art convolutional neural network architecture for object detection was used to detect operating hands in open surgery videos. Automated assessment was expanded by combining model predictions with a fast object tracker to enable surgeon-specific hand tracking. To train our model, we used publicly available videos of open surgery from YouTube and annotated these with spatial bounding boxes of operating hands. Our model's spatial detections of operating hands significantly outperforms the detections achieved using pre-existing hand-detection datasets, and allow for insights into intra-operative movement patterns and economy of motion.

Joy Hsu, Jeffrey Gu, Gong-Her Wu et al.

We consider the task of representation learning for unsupervised segmentation of 3D voxel-grid biomedical images. We show that models that capture implicit hierarchical relationships between subvolumes are better suited for this task. To that end, we consider encoder-decoder architectures with a hyperbolic latent space, to explicitly capture hierarchical relationships present in subvolumes of the data. We propose utilizing a 3D hyperbolic variational autoencoder with a novel gyroplane convolutional layer to map from the embedding space back to 3D images. To capture these relationships, we introduce an essential self-supervised loss -- in addition to the standard VAE loss -- which infers approximate hierarchies and encourages implicitly related subvolumes to be mapped closer in the embedding space. We present experiments on both synthetic data and biomedical data to validate our hypothesis.

Zhenzhen Weng, Serena Yeung

The 3D world limits the human body pose and the human body pose conveys information about the surrounding objects. Indeed, from a single image of a person placed in an indoor scene, we as humans are adept at resolving ambiguities of the human pose and room layout through our knowledge of the physical laws and prior perception of the plausible object and human poses. However, few computer vision models fully leverage this fact. In this work, we propose an end-to-end trainable model that perceives the 3D scene from a single RGB image, estimates the camera pose and the room layout, and reconstructs both human body and object meshes. By imposing a set of comprehensive and sophisticated losses on all aspects of the estimations, we show that our model outperforms existing human body mesh methods and indoor scene reconstruction methods. To the best of our knowledge, this is the first model that outputs both object and human predictions at the mesh level, and performs joint optimization on the scene and human poses.

Ali Mottaghi, Prathusha K Sarma, Xavier Amatriain et al.

We study the problem of medical symptoms recognition from patient text, for the purposes of gathering pertinent information from the patient (known as history-taking). A typical patient text is often descriptive of the symptoms the patient is experiencing and a single instance of such a text can be "labeled" with multiple symptoms. This makes learning a medical symptoms recognizer challenging on account of i) the lack of availability of voluminous annotated data as well as ii) the large unknown universe of multiple symptoms that a single text can map to. Furthermore, patient text is often characterized by a long tail in the data (i.e., some labels/symptoms occur more frequently than others for e.g "fever" vs "hematochezia"). In this paper, we introduce an active learning method that leverages underlying structure of a continually refined, learned latent space to select the most informative examples to label. This enables the selection of the most informative examples that progressively increases the coverage on the universe of symptoms via the learned model, despite the long tail in data distribution.

Sabina Tomkins, Peng Liao, Predrag Klasnja et al.

In mobile health (mHealth), reinforcement learning algorithms that adapt to one's context without learning personalized policies might fail to distinguish between the needs of individuals. Yet the high amount of noise due to the in situ delivery of mHealth interventions can cripple the ability of an algorithm to learn when given access to only a single user's data, making personalization challenging. We present IntelligentPooling, which learns personalized policies via an adaptive, principled use of other users' data. We show that IntelligentPooling achieves an average of 26% lower regret than state-of-the-art across all generative models. Additionally, we inspect the behavior of this approach in a live clinical trial, demonstrating its ability to learn from even a small group of users.

Ali Mottaghi, Serena Yeung

Active learning aims to develop label-efficient algorithms by querying the most informative samples to be labeled by an oracle. The design of efficient training methods that require fewer labels is an important research direction that allows more effective use of computational and human resources for labeling and training deep neural networks. In this work, we demonstrate how we can use recent advances in deep generative models, to outperform the state-of-the-art in achieving the highest classification accuracy using as few labels as possible. Unlike previous approaches, our approach uses not only labeled images to train the classifier but also unlabeled images and generated images for co-training the whole model. Our experiments show that the proposed method significantly outperforms existing approaches in active learning on a wide range of datasets (MNIST, CIFAR-10, SVHN, CelebA, and ImageNet).

Edward Chou, Josh Beal, Daniel Levy et al.

Homomorphic encryption enables arbitrary computation over data while it remains encrypted. This privacy-preserving feature is attractive for machine learning, but requires significant computational time due to the large overhead of the encryption scheme. We present Faster CryptoNets, a method for efficient encrypted inference using neural networks. We develop a pruning and quantization approach that leverages sparse representations in the underlying cryptosystem to accelerate inference. We derive an optimal approximation for popular activation functions that achieves maximally-sparse encodings and minimizes approximation error. We also show how privacy-safe training techniques can be used to reduce the overhead of encrypted inference for real-world datasets by leveraging transfer learning and differential privacy. Our experiments show that our method maintains competitive accuracy and achieves a significant speedup over previous methods. This work increases the viability of deep learning systems that use homomorphic encryption to protect user privacy.

Amy Jin, Serena Yeung, Jeffrey Jopling et al.

Five billion people in the world lack access to quality surgical care. Surgeon skill varies dramatically, and many surgical patients suffer complications and avoidable harm. Improving surgical training and feedback would help to reduce the rate of complications, half of which have been shown to be preventable. To do this, it is essential to assess operative skill, a process that currently requires experts and is manual, time consuming, and subjective. In this work, we introduce an approach to automatically assess surgeon performance by tracking and analyzing tool movements in surgical videos, leveraging region-based convolutional neural networks. In order to study this problem, we also introduce a new dataset, m2cai16-tool-locations, which extends the m2cai16-tool dataset with spatial bounds of tools. While previous methods have addressed tool presence detection, ours is the first to not only detect presence but also spatially localize surgical tools in real-world laparoscopic surgical videos. We show that our method both effectively detects the spatial bounds of tools as well as significantly outperforms existing methods on tool presence detection. We further demonstrate the ability of our method to assess surgical quality through analysis of tool usage patterns, movement range, and economy of motion.

Albert Haque, Michelle Guo, Alexandre Alahi et al.

One in twenty-five patients admitted to a hospital will suffer from a hospital acquired infection. If we can intelligently track healthcare staff, patients, and visitors, we can better understand the sources of such infections. We envision a smart hospital capable of increasing operational efficiency and improving patient care with less spending. In this paper, we propose a non-intrusive vision-based system for tracking people's activity in hospitals. We evaluate our method for the problem of measuring hand hygiene compliance. Empirically, our method outperforms existing solutions such as proximity-based techniques and covert in-person observational studies. We present intuitive, qualitative results that analyze human movement patterns and conduct spatial analytics which convey our method's interpretability. This work is a step towards a computer-vision based smart hospital and demonstrates promising results for reducing hospital acquired infections.

Serena Yeung, Vignesh Ramanathan, Olga Russakovsky et al.

Understanding the simultaneously very diverse and intricately fine-grained set of possible human actions is a critical open problem in computer vision. Manually labeling training videos is feasible for some action classes but doesn't scale to the full long-tailed distribution of actions. A promising way to address this is to leverage noisy data from web queries to learn new actions, using semi-supervised or "webly-supervised" approaches. However, these methods typically do not learn domain-specific knowledge, or rely on iterative hand-tuned data labeling policies. In this work, we instead propose a reinforcement learning-based formulation for selecting the right examples for training a classifier from noisy web search results. Our method uses Q-learning to learn a data labeling policy on a small labeled training dataset, and then uses this to automatically label noisy web data for new visual concepts. Experiments on the challenging Sports-1M action recognition benchmark as well as on additional fine-grained and newly emerging action classes demonstrate that our method is able to learn good labeling policies for noisy data and use this to learn accurate visual concept classifiers.

Serena Yeung, Anitha Kannan, Yann Dauphin et al.

Variational autoencoders (VAE) are directed generative models that learn factorial latent variables. As noted by Burda et al. (2015), these models exhibit the problem of factor over-pruning where a significant number of stochastic factors fail to learn anything and become inactive. This can limit their modeling power and their ability to learn diverse and meaningful latent representations. In this paper, we evaluate several methods to address this problem and propose a more effective model-based approach called the epitomic variational autoencoder (eVAE). The so-called epitomes of this model are groups of mutually exclusive latent factors that compete to explain the data. This approach helps prevent inactive units since each group is pressured to explain the data. We compare the approaches with qualitative and quantitative results on MNIST and TFD datasets. Our results show that eVAE makes efficient use of model capacity and generalizes better than VAE.

Albert Haque, Boya Peng, Zelun Luo et al.

We propose a viewpoint invariant model for 3D human pose estimation from a single depth image. To achieve this, our discriminative model embeds local regions into a learned viewpoint invariant feature space. Formulated as a multi-task learning problem, our model is able to selectively predict partial poses in the presence of noise and occlusion. Our approach leverages a convolutional and recurrent network architecture with a top-down error feedback mechanism to self-correct previous pose estimates in an end-to-end manner. We evaluate our model on a previously published depth dataset and a newly collected human pose dataset containing 100K annotated depth images from extreme viewpoints. Experiments show that our model achieves competitive performance on frontal views while achieving state-of-the-art performance on alternate viewpoints.

Serena Yeung, Olga Russakovsky, Greg Mori et al.

In this work we introduce a fully end-to-end approach for action detection in videos that learns to directly predict the temporal bounds of actions. Our intuition is that the process of detecting actions is naturally one of observation and refinement: observing moments in video, and refining hypotheses about when an action is occurring. Based on this insight, we formulate our model as a recurrent neural network-based agent that interacts with a video over time. The agent observes video frames and decides both where to look next and when to emit a prediction. Since backpropagation is not adequate in this non-differentiable setting, we use REINFORCE to learn the agent's decision policy. Our model achieves state-of-the-art results on the THUMOS'14 and ActivityNet datasets while observing only a fraction (2% or less) of the video frames.