Derek Hoiem

Zhenhailong Wang, Manling Li, Ruochen Xu et al.

The goal of this work is to build flexible video-language models that can generalize to various video-to-text tasks from few examples, such as domain-specific captioning, question answering, and future event prediction. Existing few-shot video-language learners focus exclusively on the encoder, resulting in the absence of a video-to-text decoder to handle generative tasks. Video captioners have been pretrained on large-scale video-language datasets, but they rely heavily on finetuning and lack the ability to generate text for unseen tasks in a few-shot setting. We propose VidIL, a few-shot Video-language Learner via Image and Language models, which demonstrates strong performance on few-shot video-to-text tasks without the necessity of pretraining or finetuning on any video datasets. We use the image-language models to translate the video content into frame captions, object, attribute, and event phrases, and compose them into a temporal structure template. We then instruct a language model, with a prompt containing a few in-context examples, to generate a target output from the composed content. The flexibility of prompting allows the model to capture any form of text input, such as automatic speech recognition (ASR) transcripts. Our experiments demonstrate the power of language models in understanding videos on a wide variety of video-language tasks, including video captioning, video question answering, video caption retrieval, and video future event prediction. Especially, on video future event prediction, our few-shot model significantly outperforms state-of-the-art supervised models trained on large-scale video datasets. Code and resources are publicly available for research purposes at https://github.com/MikeWangWZHL/VidIL .

Yangyi Chen, Xingyao Wang, Manling Li et al.

State-of-the-art vision-language models (VLMs) still have limited performance in structural knowledge extraction, such as relations between objects. In this work, we present ViStruct, a training framework to learn VLMs for effective visual structural knowledge extraction. Two novel designs are incorporated. First, we propose to leverage the inherent structure of programming language to depict visual structural information. This approach enables explicit and consistent representation of visual structural information of multiple granularities, such as concepts, relations, and events, in a well-organized structured format. Second, we introduce curriculum-based learning for VLMs to progressively comprehend visual structures, from fundamental visual concepts to intricate event structures. Our intuition is that lower-level knowledge may contribute to complex visual structure understanding. Furthermore, we compile and release a collection of datasets tailored for visual structural knowledge extraction. We adopt a weakly-supervised approach to directly generate visual event structures from captions for ViStruct training, capitalizing on abundant image-caption pairs from the web. In experiments, we evaluate ViStruct on visual structure prediction tasks, demonstrating its effectiveness in improving the understanding of visual structures. The code is public at \url{https://github.com/Yangyi-Chen/vi-struct}.

Zhen Zhu, Yiming Gong, Derek Hoiem

We propose an approach for anytime continual learning (AnytimeCL) for open vocabulary image classification. The AnytimeCL problem aims to break away from batch training and rigid models by requiring that a system can predict any set of labels at any time and efficiently update and improve when receiving one or more training samples at any time. Despite the challenging goal, we achieve substantial improvements over recent methods. We propose a dynamic weighting between predictions of a partially fine-tuned model and a fixed open vocabulary model that enables continual improvement when training samples are available for a subset of a task's labels. We also propose an attention-weighted PCA compression of training features that reduces storage and computation with little impact to model accuracy. Our methods are validated with experiments that test flexibility of learning and inference. Code is available at https://github.com/jessemelpolio/AnytimeCL.

Zhen Zhu, Weijie Lyu, Yao Xiao et al.

We introduce a method for flexible and efficient continual learning in open-vocabulary image classification, drawing inspiration from the complementary learning systems observed in human cognition. Specifically, we propose to combine predictions from a CLIP zero-shot model and the exemplar-based model, using the zero-shot estimated probability that a sample's class is within the exemplar classes. We also propose a "tree probe" method, an adaption of lazy learning principles, which enables fast learning from new examples with competitive accuracy to batch-trained linear models. We test in data incremental, class incremental, and task incremental settings, as well as ability to perform flexible inference on varying subsets of zero-shot and learned categories. Our proposed method achieves a good balance of learning speed, target task effectiveness, and zero-shot effectiveness. Code will be available at https://github.com/jessemelpolio/TreeProbe.

Zhen Zhu, Yijun Li, Weijie Lyu et al.

We investigate how to generate multimodal image outputs, such as RGB, depth, and surface normals, with a single generative model. The challenge is to produce outputs that are realistic, and also consistent with each other. Our solution builds on the StyleGAN3 architecture, with a shared backbone and modality-specific branches in the last layers of the synthesis network, and we propose per-modality fidelity discriminators and a cross-modality consistency discriminator. In experiments on the Stanford2D3D dataset, we demonstrate realistic and consistent generation of RGB, depth, and normal images. We also show a training recipe to easily extend our pretrained model on a new domain, even with a few pairwise data. We further evaluate the use of synthetically generated RGB and depth pairs for training or fine-tuning depth estimators. Code will be available at https://github.com/jessemelpolio/MultimodalGAN.

Anand Bhattad, Daniel McKee, Derek Hoiem et al.

Intrinsic images, in the original sense, are image-like maps of scene properties like depth, normal, albedo or shading. This paper demonstrates that StyleGAN can easily be induced to produce intrinsic images. The procedure is straightforward. We show that, if StyleGAN produces $G({w})$ from latents ${w}$, then for each type of intrinsic image, there is a fixed offset ${d}_c$ so that $G({w}+{d}_c)$ is that type of intrinsic image for $G({w})$. Here ${d}_c$ is {\em independent of ${w}$}. The StyleGAN we used was pretrained by others, so this property is not some accident of our training regime. We show that there are image transformations StyleGAN will {\em not} produce in this fashion, so StyleGAN is not a generic image regression engine. It is conceptually exciting that an image generator should ``know'' and represent intrinsic images. There may also be practical advantages to using a generative model to produce intrinsic images. The intrinsic images obtained from StyleGAN compare well both qualitatively and quantitatively with those obtained by using SOTA image regression techniques; but StyleGAN's intrinsic images are robust to relighting effects, unlike SOTA methods.

Tanmay Gupta, Ryan Marten, Aniruddha Kembhavi et al.

Computer vision models excel at making predictions when the test distribution closely resembles the training distribution. Such models have yet to match the ability of biological vision to learn from multiple sources and generalize to new data sources and tasks. To facilitate the development and evaluation of more general vision systems, we introduce the General Robust Image Task (GRIT) benchmark. GRIT evaluates the performance, robustness, and calibration of a vision system across a variety of image prediction tasks, concepts, and data sources. The seven tasks in GRIT are selected to cover a range of visual skills: object categorization, object localization, referring expression grounding, visual question answering, segmentation, human keypoint detection, and surface normal estimation. GRIT is carefully designed to enable the evaluation of robustness under image perturbations, image source distribution shift, and concept distribution shift. By providing a unified platform for thorough assessment of skills and concepts learned by a vision model, we hope GRIT catalyzes the development of performant and robust general purpose vision systems.

Anand Bhattad, Viraj Shah, Derek Hoiem et al.

StyleGAN's disentangled style representation enables powerful image editing by manipulating the latent variables, but accurately mapping real-world images to their latent variables (GAN inversion) remains a challenge. Existing GAN inversion methods struggle to maintain editing directions and produce realistic results. To address these limitations, we propose Make It So, a novel GAN inversion method that operates in the $\mathcal{Z}$ (noise) space rather than the typical $\mathcal{W}$ (latent style) space. Make It So preserves editing capabilities, even for out-of-domain images. This is a crucial property that was overlooked in prior methods. Our quantitative evaluations demonstrate that Make It So outperforms the state-of-the-art method PTI~\cite{roich2021pivotal} by a factor of five in inversion accuracy and achieves ten times better edit quality for complex indoor scenes.

Michal Shlapentokh-Rothman, Prachi Garg, Yu-Xiong Wang et al.

Keyframe selection is a direct way to provide verifiable visual evidence for long-video question answering (QA). Queries differ in what they require, and finding the right frames depends on knowing what to look for. Existing keyframe selectors either score every frame against a single query, or decompose the query into a fixed schema evaluated by a single visual tool. We propose ToolMerge, a keyframe retrieval method based on decomposition and merging: an Large Language Model (LLM) based planner decomposes the query into tool calls and specifies how their per-tool rankings are merged using boolean operators. To evaluate retrieval directly, we construct Molmo-2 Moments (M2M), a benchmark in which every question is anchored to a specific time interval by construction. Across QA, question retrieval, and caption retrieval, ToolMerge is competitive with prior keyframe selectors, most notably on caption retrieval, outperforming other methods by 5%. Code and data can be found at https://github.com/michalsr/ToolMerge .

Savya Khosla, Sethuraman T, Aryan Chadha et al.

Despite recent progress, vision-language encoders struggle with two core limitations: (1) weak alignment between language and dense vision features, which hurts tasks like open-vocabulary semantic segmentation; and (2) high token counts for fine-grained visual representations, which limits scalability to long videos. This work addresses both limitations. We propose T-REN (Text-aligned Region Encoder Network), an efficient encoder that maps visual data to a compact set of text-aligned region-level representations (or region tokens). T-REN achieves this through a lightweight network added on top of a frozen vision backbone, trained to pool patch-level representations within each semantic region into region tokens and align them with region-level text annotations. With only 3.7% additional parameters compared to the vision-language backbone, this design yields substantially stronger dense cross-modal understanding while reducing the token count by orders of magnitude. Specifically, T-REN delivers +5.9 mIoU on ADE20K open-vocabulary segmentation, +18.4% recall on COCO object-level text-image retrieval, +15.6% recall on Ego4D video object localization, and +17.6% mIoU on VSPW video scene parsing, all while reducing token counts by more than 24x for images and 187x for videos compared to the patch-based vision-language backbone. The code and model are available at https://github.com/savya08/T-REN.

Jae Yong Lee, Yuqun Wu, Chuhang Zou et al.

Multilayer perceptrons (MLPs) learn high frequencies slowly. Recent approaches encode features in spatial bins to improve speed of learning details, but at the cost of larger model size and loss of continuity. Instead, we propose to encode features in bins of Fourier features that are commonly used for positional encoding. We call these Quantized Fourier Features (QFF). As a naturally multiresolution and periodic representation, our experiments show that using QFF can result in smaller model size, faster training, and better quality outputs for several applications, including Neural Image Representations (NIR), Neural Radiance Field (NeRF) and Signed Distance Function (SDF) modeling. QFF are easy to code, fast to compute, and serve as a simple drop-in addition to many neural field representations.

Yuqun Wu, Jae Yong Lee, Derek Hoiem

Our long term goal is to use image-based depth completion to quickly create 3D models from sparse point clouds, e.g. from SfM or SLAM. Much progress has been made in depth completion. However, most current works assume well distributed samples of known depth, e.g. Lidar or random uniform sampling, and perform poorly on uneven samples, such as from keypoints, due to the large unsampled regions. To address this problem, we extend CSPN with multiscale prediction and a dilated kernel, leading to much better completion of keypoint-sampled depth. We also show that a model trained on NYUv2 creates surprisingly good point clouds on ETH3D by completing sparse SfM points.

Jae Yong Lee, Chuhang Zou, Derek Hoiem

Recent work in multi-view stereo (MVS) combines learnable photometric scores and regularization with PatchMatch-based optimization to achieve robust pixelwise estimates of depth, normals, and visibility. However, non-learning based methods still outperform for large scenes with sparse views, in part due to use of geometric consistency constraints and ability to optimize over many views at high resolution. In this paper, we build on learning-based approaches to improve photometric scores by learning patch coplanarity and encourage geometric consistency by learning a scaled photometric cost that can be combined with reprojection error. We also propose an adaptive pixel sampling strategy for candidate propagation that reduces memory to enable training on larger resolution with more views and a larger encoder. These modifications lead to 6-15% gains in accuracy and completeness on the challenging ETH3D benchmark, resulting in higher F1 performance than the widely used state-of-the-art non-learning approaches ACMM and ACMP.

Guy Kaplan, Zorik Gekhman, Zhen Zhu et al.

Large language models are prone to hallucinating factually incorrect statements. A key source of these errors is exposure to new factual information through supervised fine-tuning (SFT), which can increase hallucinations w.r.t. knowledge acquired during pre-training. In this work, we explore whether SFT-induced hallucinations can be mitigated using established tools from the continual learning literature, since they arise as a by-product of knowledge degradation during training. We propose a self-distillation-based SFT method that facilitates effective factual learning while minimizing hallucinations w.r.t. pre-existing knowledge by regularizing output-distribution drift. We also show that, in settings where new knowledge acquisition is unnecessary, suppressing factual plasticity by freezing parameter groups, can preserve task performance while reducing hallucinations. Lastly, we investigate the mechanism behind SFT-induced hallucinations through three hypotheses: capacity limitations, behavior cloning, and localized interference. Our experiments show that a main driver is interference among overlapping semantic representations, and that self-distillation succeeds by mitigating this interference.

Heyi Tao, Sethuraman T, Michal Shlapentokh-Rothman et al.

The paper investigates using a Large Language Model (LLM) to automatically perform web software tasks using click, scroll, and text input operations. Previous approaches, such as reinforcement learning (RL) or imitation learning, are inefficient to train and task-specific. Our method uses filtered Document Object Model (DOM) elements as observations and performs tasks step-by-step, sequentially generating small programs based on the current observations. We use in-context learning, either benefiting from a single manually provided example, or an automatically generated example based on a successful zero-shot trial. We evaluate the proposed method on the MiniWob++ benchmark. With only one in-context example, our WebWISE method achieves similar or better performance than other methods that require many demonstrations or trials.

Jae Yong Lee, Yuqun Wu, Chuhang Zou et al.

The goal of this paper is to encode a 3D scene into an extremely compact representation from 2D images and to enable its transmittance, decoding and rendering in real-time across various platforms. Despite the progress in NeRFs and Gaussian Splats, their large model size and specialized renderers make it challenging to distribute free-viewpoint 3D content as easily as images. To address this, we have designed a novel 3D representation that encodes the plenoptic function into sinusoidal function indexed dense volumes. This approach facilitates feature sharing across different locations, improving compactness over traditional spatial voxels. The memory footprint of the dense 3D feature grid can be further reduced using spatial decomposition techniques. This design combines the strengths of spatial hashing functions and voxel decomposition, resulting in a model size as small as 150 KB for each 3D scene. Moreover, PPNG features a lightweight rendering pipeline with only 300 lines of code that decodes its representation into standard GL textures and fragment shaders. This enables real-time rendering using the traditional GL pipeline, ensuring universal compatibility and efficiency across various platforms without additional dependencies.

Yuqun Wu, Chih-hao Lin, Henry Che et al.

We investigate the problem of identifying objects that have been added, removed, or moved between a pair of captures (images or videos) of the same scene at different times. Detecting such changes is important for many applications, such as robotic tidying or construction progress and safety monitoring. A major challenge is that varying viewpoints can cause objects to falsely appear changed. We introduce SceneDiff Benchmark, the first multiview change detection benchmark with object instance annotations, comprising 350 diverse video pairs with thousands of changed objects. We also introduce the SceneDiff method, a new training-free approach for multiview object change detection that leverages pretrained 3D, segmentation, and image encoding models to robustly predict across multiple benchmarks. Our method aligns the captures in 3D, extracts object regions, and compares spatial and semantic region features to detect changes. Experiments on multi-view and two-view benchmarks demonstrate that our method outperforms existing approaches by large margins (94% and 37.4% relative AP improvements). The benchmark and code will be publicly released.

Jiasen Lu, Christopher Clark, Sangho Lee et al. · allen-ai

We present Unified-IO 2, the first autoregressive multimodal model that is capable of understanding and generating image, text, audio, and action. To unify different modalities, we tokenize inputs and outputs -- images, text, audio, action, bounding boxes, etc., into a shared semantic space and then process them with a single encoder-decoder transformer model. Since training with such diverse modalities is challenging, we propose various architectural improvements to stabilize model training. We train our model from scratch on a large multimodal pre-training corpus from diverse sources with a multimodal mixture of denoisers objective. To learn an expansive set of skills, such as following multimodal instructions, we construct and finetune on an ensemble of 120 datasets with prompts and augmentations. With a single unified model, Unified-IO 2 achieves state-of-the-art performance on the GRIT benchmark and strong results in more than 35 benchmarks, including image generation and understanding, natural language understanding, video and audio understanding, and robotic manipulation. We release all our models to the research community.

Zhen Zhu, Yiming Gong, Yao Xiao et al.

How can we teach large multimodal models (LMMs) new skills without erasing prior abilities? We study sequential fine-tuning on five target skills while monitoring general ability on eight held-out benchmarks across three model families. We observe that apparent "forgetting" on held-out tasks after narrow fine-tuning can partly recover at later stages. We trace this behavior to a measurable shift in the output token distribution, manifested through a simple counting-bias probe that co-varies with forgetting. Guided by this picture, we identify two simple, robust tuning recipes that learn strongly while limiting drift: (i) updating only the self-attention projection layers, and (ii) updating only the MLP Gate&Up while freezing the Down projection. Across models and tasks, these choices deliver strong target gains while largely preserving held-out performance. Code is available at https://github.com/jessemelpolio/LMM_CL

Yao Xiao, Qiqian Fu, Heyi Tao et al.

Image-text models excel at image-level tasks but struggle with detailed visual understanding. While these models provide strong visual-language alignment, segmentation models like SAM2 offer precise spatial boundaries for objects. To this end, we propose TextRegion, a simple, effective, and training-free framework that combines the strengths of image-text models and SAM2 to generate powerful text-aligned region tokens. These tokens enable detailed visual understanding while preserving open-vocabulary capabilities. They can be directly applied to various downstream tasks, including open-world semantic segmentation, referring expression comprehension, and grounding. We conduct extensive evaluations and consistently achieve superior or competitive performance compared to state-of-the-art training-free methods. Additionally, our framework is compatible with many image-text models, making it highly practical and easily extensible as stronger models emerge. Code is available at: https://github.com/avaxiao/TextRegion.

Savya Khosla, Sethuraman TV, Barnett Lee et al.

We introduce the Region Encoder Network (REN), a fast and effective model for generating region-based image representations using point prompts. Recent methods combine class-agnostic segmenters (e.g., SAM) with patch-based image encoders (e.g., DINO) to produce compact and effective region representations, but they suffer from high computational cost due to the segmentation step. REN bypasses this bottleneck using a lightweight module that directly generates region tokens, enabling 60x faster token generation with 35x less memory, while also improving token quality. It uses a few cross-attention blocks that take point prompts as queries and features from a patch-based image encoder as keys and values to produce region tokens that correspond to the prompted objects. We train REN with three popular encoders-DINO, DINOv2, and OpenCLIP-and show that it can be extended to other encoders without dedicated training. We evaluate REN on semantic segmentation and retrieval tasks, where it consistently outperforms the original encoders in both performance and compactness, and matches or exceeds SAM-based region methods while being significantly faster. Notably, REN achieves state-of-the-art results on the challenging Ego4D VQ2D benchmark and outperforms proprietary LMMs on Visual Haystacks' single-needle challenge. Code and models are available at: https://github.com/savya08/REN.

Hongxu Yin, Pavlo Molchanov, Zhizhong Li et al.

We introduce DeepInversion, a new method for synthesizing images from the image distribution used to train a deep neural network. We 'invert' a trained network (teacher) to synthesize class-conditional input images starting from random noise, without using any additional information about the training dataset. Keeping the teacher fixed, our method optimizes the input while regularizing the distribution of intermediate feature maps using information stored in the batch normalization layers of the teacher. Further, we improve the diversity of synthesized images using Adaptive DeepInversion, which maximizes the Jensen-Shannon divergence between the teacher and student network logits. The resulting synthesized images from networks trained on the CIFAR-10 and ImageNet datasets demonstrate high fidelity and degree of realism, and help enable a new breed of data-free applications - ones that do not require any real images or labeled data. We demonstrate the applicability of our proposed method to three tasks of immense practical importance -- (i) data-free network pruning, (ii) data-free knowledge transfer, and (iii) data-free continual learning. Code is available at https://github.com/NVlabs/DeepInversion

Zhizhong Li, Derek Hoiem

Intuitively, unfamiliarity should lead to lack of confidence. In reality, current algorithms often make highly confident yet wrong predictions when faced with relevant but unfamiliar examples. A classifier we trained to recognize gender is 12 times more likely to be wrong with a 99% confident prediction if presented with a subject from a different age group than those seen during training. In this paper, we compare and evaluate several methods to improve confidence estimates for unfamiliar and familiar samples. We propose a testing methodology of splitting unfamiliar and familiar samples by attribute (age, breed, subcategory) or sampling (similar datasets collected by different people at different times). We evaluate methods including confidence calibration, ensembles, distillation, and a Bayesian model and use several metrics to analyze label, likelihood, and calibration error. While all methods reduce over-confident errors, the ensemble of calibrated models performs best overall, and T-scaling performs best among the approaches with fastest inference. Our code is available at https://github.com/lizhitwo/ConfidenceEstimates . $\color{red}{\text{Please see UPDATED ERRATA.}}$

Michal Shlapentokh-Rothman, Ansel Blume, Yao Xiao et al.

We investigate whether region-based representations are effective for recognition. Regions were once a mainstay in recognition approaches, but pixel and patch-based features are now used almost exclusively. We show that recent class-agnostic segmenters like SAM can be effectively combined with strong unsupervised representations like DINOv2 and used for a wide variety of tasks, including semantic segmentation, object-based image retrieval, and multi-image analysis. Once the masks and features are extracted, these representations, even with linear decoders, enable competitive performance, making them well suited to applications that require custom queries. The compactness of the representation also makes it well-suited to video analysis and other problems requiring inference across many images.

Savya Khosla, Sethuraman T, Alexander Schwing et al.

We present RELOCATE, a simple training-free baseline designed to perform the challenging task of visual query localization in long videos. To eliminate the need for task-specific training and efficiently handle long videos, RELOCATE leverages a region-based representation derived from pretrained vision models. At a high level, it follows the classic object localization approach: (1) identify all objects in each video frame, (2) compare the objects with the given query and select the most similar ones, and (3) perform bidirectional tracking to get a spatio-temporal response. However, we propose some key enhancements to handle small objects, cluttered scenes, partial visibility, and varying appearances. Notably, we refine the selected objects for accurate localization and generate additional visual queries to capture visual variations. We evaluate RELOCATE on the challenging Ego4D Visual Query 2D Localization dataset, establishing a new baseline that outperforms prior task-specific methods by 49% (relative improvement) in spatio-temporal average precision.

Sethuraman TV, Savya Khosla, Vignesh Srinivasakumar et al.

Dense video prediction tasks, such as object tracking and semantic segmentation, require video encoders that generate temporally consistent, spatially dense features for every frame. However, existing approaches fall short: image encoders like DINO or CLIP lack temporal awareness, while video models such as VideoMAE underperform compared to image encoders on dense prediction tasks. We address this gap with FRAME, a self-supervised video frame encoder tailored for dense video understanding. FRAME learns to predict current and future DINO patch features from past and present RGB frames, leading to spatially precise and temporally coherent representations. To our knowledge, FRAME is the first video encoder to leverage image-based models for dense prediction while outperforming them on tasks requiring fine-grained visual correspondence. As an auxiliary capability, FRAME aligns its class token with CLIP's semantic space, supporting language-driven tasks such as video classification. We evaluate FRAME across six dense prediction tasks on seven datasets, where it consistently outperforms image encoders and existing self-supervised video models. Despite its versatility, FRAME maintains a compact architecture suitable for a range of downstream applications.

Yuqun Wu, Jae Yong Lee, Chuhang Zou et al.

The latest regularized Neural Radiance Field (NeRF) approaches produce poor geometry and view extrapolation for large scale sparse view scenes, such as ETH3D. Density-based approaches tend to be under-constrained, while surface-based approaches tend to miss details. In this paper, we take a density-based approach, sampling patches instead of individual rays to better incorporate monocular depth and normal estimates and patch-based photometric consistency constraints between training views and sampled virtual views. Loosely constraining densities based on estimated depth aligned to sparse points further improves geometric accuracy. While maintaining similar view synthesis quality, our approach significantly improves geometric accuracy on the ETH3D benchmark, e.g. increasing the F1@2cm score by 4x-8x compared to other regularized density-based approaches, with much lower training and inference time than other approaches.

Sethuraman T, Savya Khosla, Aditi Tiwari et al.

This work investigates a fundamental question: Do Video-Language Models (VidLMs) robustly account for video content, temporal sequence, and motion? Our investigation shows that, surprisingly, they often do not. We introduce REVEAL{}, a diagnostic benchmark that probes fundamental weaknesses of contemporary VidLMs through five controlled stress tests; assessing temporal expectation bias, reliance on language-only shortcuts, video sycophancy, camera motion sensitivity, and robustness to spatiotemporal occlusion. We test leading open- and closed-source VidLMs and find that these models confidently describe reversed scenes as forward, answer questions while neglecting video content, agree with false claims, struggle with basic camera motion, and fail to aggregate temporal information amidst simple spatiotemporal masking. Humans, on the other hand, succeed at these tasks with ease. Alongside our benchmark, we provide a data pipeline that automatically generates diagnostic examples for our stress tests, enabling broader and more scalable evaluation. We will release our benchmark and code to support future research.

Ansel Blume, Jeonghwan Kim, Hyeonjeong Ha et al.

Real-world objects are composed of distinctive, object-specific parts. Identifying these parts is key to performing fine-grained, compositional reasoning-yet, large multimodal models (LMMs) struggle to perform this seemingly straightforward task. In this work, we introduce PARTONOMY, an LMM benchmark designed for pixel-level part grounding. We construct PARTONOMY from existing part datasets and our own rigorously annotated set of images, encompassing 862 part labels and 534 object labels for evaluation. Unlike existing datasets that simply ask models to identify generic parts, PARTONOMY uses specialized concepts (e.g., agricultural airplane), and challenges models to compare objects' parts, consider part-whole relationships, and justify textual predictions with visual segmentations. Our experiments demonstrate significant limitations in state-of-the-art LMMs (e.g., LISA-13B achieves only 5.9% gIoU), highlighting a critical gap in their part grounding abilities. We note that existing segmentation-enabled LMMs (segmenting LMMs) have two key architectural shortcomings: they use special [SEG] tokens not seen during pretraining which induce distribution shift, and they discard predicted segmentations instead of using past predictions to guide future ones. To address these deficiencies, we train several part-centric LMMs and propose PLUM, a novel segmenting LMM that uses span tagging instead of segmentation tokens and that conditions on prior predictions in a feedback loop. We find that pretrained PLUM outperforms existing segmenting LMMs on reasoning segmentation, VQA, and visual hallucination benchmarks. In addition, PLUM finetuned on our proposed Explanatory Part Segmentation task is competitive with segmenting LMMs trained on significantly more segmentation data. Our work opens up new avenues towards enabling fine-grained, grounded visual understanding in LMMs.

Michal Shlapentokh-Rothman, Yu-Xiong Wang, Derek Hoiem

Adapting visual programming or prompting large language models (LLMs) to generate executable code for visual tasks like visual question answering (VQA) for specialized tasks or domains remains challenging due to high annotation and inference costs. We propose a low-cost visual program distillation method that can be used for models with at most 1 billion parameters and requires no human-generated program annotations. We achieve this through synthetic data augmentation based on decoupling programs into higher-level skills, called templates, and their corresponding arguments. Experimental results show that, with a relatively small amount of question/answer data, small language models can generate high-quality specialized visual programs with the added benefit of much faster inference

Amita Kamath, Christopher Clark, Tanmay Gupta et al.

General Purpose Vision (GPV) systems are models that are designed to solve a wide array of visual tasks without requiring architectural changes. Today, GPVs primarily learn both skills and concepts from large fully supervised datasets. Scaling GPVs to tens of thousands of concepts by acquiring data to learn each concept for every skill quickly becomes prohibitive. This work presents an effective and inexpensive alternative: learn skills from supervised datasets, learn concepts from web image search, and leverage a key characteristic of GPVs: the ability to transfer visual knowledge across skills. We use a dataset of 1M+ images spanning 10k+ visual concepts to demonstrate webly-supervised concept expansion for two existing GPVs (GPV-1 and VL-T5) on 3 benchmarks: 5 COCO-based datasets (80 primary concepts), a newly curated series of 5 datasets based on the OpenImages and VisualGenome repositories (~500 concepts), and the Web-derived dataset (10k+ concepts). We also propose a new architecture, GPV-2 that supports a variety of tasks -- from vision tasks like classification and localization to vision+language tasks like QA and captioning, to more niche ones like human-object interaction detection. GPV-2 benefits hugely from web data and outperforms GPV-1 and VL-T5 across these benchmarks. Our data, code, and web demo are available at https://prior.allenai.org/projects/gpv2.

Jae Yong Lee, Joseph DeGol, Chuhang Zou et al.

Recent learning-based multi-view stereo (MVS) methods show excellent performance with dense cameras and small depth ranges. However, non-learning based approaches still outperform for scenes with large depth ranges and sparser wide-baseline views, in part due to their PatchMatch optimization over pixelwise estimates of depth, normals, and visibility. In this paper, we propose an end-to-end trainable PatchMatch-based MVS approach that combines advantages of trainable costs and regularizations with pixelwise estimates. To overcome the challenge of the non-differentiable PatchMatch optimization that involves iterative sampling and hard decisions, we use reinforcement learning to minimize expected photometric cost and maximize likelihood of ground truth depth and normals. We incorporate normal estimation by using dilated patch kernels, and propose a recurrent cost regularization that applies beyond frontal plane-sweep algorithms to our pixelwise depth/normal estimates. We evaluate our method on widely used MVS benchmarks, ETH3D and Tanks and Temples (TnT), and compare to other state of the art learning based MVS models. On ETH3D, our method outperforms other recent learning-based approaches and performs comparably on advanced TnT.

Tanmay Gupta, Amita Kamath, Aniruddha Kembhavi et al.

Computer vision systems today are primarily N-purpose systems, designed and trained for a predefined set of tasks. Adapting such systems to new tasks is challenging and often requires non-trivial modifications to the network architecture (e.g. adding new output heads) or training process (e.g. adding new losses). To reduce the time and expertise required to develop new applications, we would like to create general purpose vision systems that can learn and perform a range of tasks without any modification to the architecture or learning process. In this paper, we propose GPV-1, a task-agnostic vision-language architecture that can learn and perform tasks that involve receiving an image and producing text and/or bounding boxes, including classification, localization, visual question answering, captioning, and more. We also propose evaluations of generality of architecture, skill-concept transfer, and learning efficiency that may inform future work on general purpose vision. Our experiments indicate GPV-1 is effective at multiple tasks, reuses some concept knowledge across tasks, can perform the Referring Expressions task zero-shot, and further improves upon the zero-shot performance using a few training samples.

Derek Hoiem, Tanmay Gupta, Zhizhong Li et al.

Learning curves model a classifier's test error as a function of the number of training samples. Prior works show that learning curves can be used to select model parameters and extrapolate performance. We investigate how to use learning curves to evaluate design choices, such as pretraining, architecture, and data augmentation. We propose a method to robustly estimate learning curves, abstract their parameters into error and data-reliance, and evaluate the effectiveness of different parameterizations. Our experiments exemplify use of learning curves for analysis and yield several interesting observations.

Tanmay Gupta, Arash Vahdat, Gal Chechik et al.

Phrase grounding, the problem of associating image regions to caption words, is a crucial component of vision-language tasks. We show that phrase grounding can be learned by optimizing word-region attention to maximize a lower bound on mutual information between images and caption words. Given pairs of images and captions, we maximize compatibility of the attention-weighted regions and the words in the corresponding caption, compared to non-corresponding pairs of images and captions. A key idea is to construct effective negative captions for learning through language model guided word substitutions. Training with our negatives yields a $\sim10\%$ absolute gain in accuracy over randomly-sampled negatives from the training data. Our weakly supervised phrase grounding model trained on COCO-Captions shows a healthy gain of $5.7\%$ to achieve $76.7\%$ accuracy on Flickr30K Entities benchmark.

Kevin Karsch, Zicheng Liao, Jason Rock et al.

Early work in computer vision considered a host of geometric cues for both shape reconstruction and recognition. However, since then, the vision community has focused heavily on shading cues for reconstruction, and moved towards data-driven approaches for recognition. In this paper, we reconsider these perhaps overlooked "boundary" cues (such as self occlusions and folds in a surface), as well as many other established constraints for shape reconstruction. In a variety of user studies and quantitative tasks, we evaluate how well these cues inform shape reconstruction (relative to each other) in terms of both shape quality and shape recognition. Our findings suggest many new directions for future research in shape reconstruction, such as automatic boundary cue detection and relaxing assumptions in shape from shading (e.g. orthographic projection, Lambertian surfaces).

Chuhang Zou, Jheng-Wei Su, Chi-Han Peng et al.

Recent approaches for predicting layouts from 360 panoramas produce excellent results. These approaches build on a common framework consisting of three steps: a pre-processing step based on edge-based alignment, prediction of layout elements, and a post-processing step by fitting a 3D layout to the layout elements. Until now, it has been difficult to compare the methods due to multiple different design decisions, such as the encoding network (e.g. SegNet or ResNet), type of elements predicted (e.g. corners, wall/floor boundaries, or semantic segmentation), or method of fitting the 3D layout. To address this challenge, we summarize and describe the common framework, the variants, and the impact of the design decisions. For a complete evaluation, we also propose extended annotations for the Matterport3D dataset [3], and introduce two depth-based evaluation metrics.

Tanmay Gupta, Alexander Schwing, Derek Hoiem

We propose to learn word embeddings from visual co-occurrences. Two words co-occur visually if both words apply to the same image or image region. Specifically, we extract four types of visual co-occurrences between object and attribute words from large-scale, textually-annotated visual databases like VisualGenome and ImageNet. We then train a multi-task log-bilinear model that compactly encodes word "meanings" represented by each co-occurrence type into a single visual word-vector. Through unsupervised clustering, supervised partitioning, and a zero-shot-like generalization analysis we show that our word embeddings complement text-only embeddings like GloVe by better representing similarities and differences between visual concepts that are difficult to obtain from text corpora alone. We further evaluate our embeddings on five downstream applications, four of which are vision-language tasks. Augmenting GloVe with our embeddings yields gains on all tasks. We also find that random embeddings perform comparably to learned embeddings on all supervised vision-language tasks, contrary to conventional wisdom.

Zhizhong Li, Linjie Luo, Sergey Tulyakov et al.

Some tasks, such as surface normals or single-view depth estimation, require per-pixel ground truth that is difficult to obtain on real images but easy to obtain on synthetic. However, models learned on synthetic images often do not generalize well to real images due to the domain shift. Our key idea to improve domain adaptation is to introduce a separate anchor task (such as facial landmarks) whose annotations can be obtained at no cost or are already available on both synthetic and real datasets. To further leverage the implicit relationship between the anchor and main tasks, we apply our \freeze technique that learns the cross-task guidance on the source domain with the final network layers, and use it on the target domain. We evaluate our methods on surface normal estimation on two pairs of datasets (indoor scenes and faces) with two kinds of anchor tasks (semantic segmentation and facial landmarks). We show that blindly applying domain adaptation or training the auxiliary task on only one domain may hurt performance, while using anchor tasks on both domains is better behaved. Our \freeze technique outperforms competing approaches, reaching performance in facial images on par with a recently popular surface normal estimation method using shape from shading domain knowledge.

Chuhang Zou, Derek Hoiem

One major challenge in 3D reconstruction is to infer the complete shape geometry from partial foreground occlusions. In this paper, we propose a method to reconstruct the complete 3D shape of an object from a single RGB image, with robustness to occlusion. Given the image and a silhouette of the visible region, our approach completes the silhouette of the occluded region and then generates a point cloud. We show improvements for reconstruction of non-occluded and partially occluded objects by providing the predicted complete silhouette as guidance. We also improve state-of-the-art for 3D shape prediction with a 2D reprojection loss from multiple synthetic views and a surface-based smoothing and refinement step. Experiments demonstrate the efficacy of our approach both quantitatively and qualitatively on synthetic and real scene datasets.

Tanmay Gupta, Alexander Schwing, Derek Hoiem

We show that for human-object interaction detection a relatively simple factorized model with appearance and layout encodings constructed from pre-trained object detectors outperforms more sophisticated approaches. Our model includes factors for detection scores, human and object appearance, and coarse (box-pair configuration) and optionally fine-grained layout (human pose). We also develop training techniques that improve learning efficiency by: (1) eliminating a train-inference mismatch; (2) rejecting easy negatives during mini-batch training; and (3) using a ratio of negatives to positives that is two orders of magnitude larger than existing approaches. We conduct a thorough ablation study to understand the importance of different factors and training techniques using the challenging HICO-Det dataset.

Daeyun Shin, Charless C. Fowlkes, Derek Hoiem

The goal of this paper is to compare surface-based and volumetric 3D object shape representations, as well as viewer-centered and object-centered reference frames for single-view 3D shape prediction. We propose a new algorithm for predicting depth maps from multiple viewpoints, with a single depth or RGB image as input. By modifying the network and the way models are evaluated, we can directly compare the merits of voxels vs. surfaces and viewer-centered vs. object-centered for familiar vs. unfamiliar objects, as predicted from RGB or depth images. Among our findings, we show that surface-based methods outperform voxel representations for objects from novel classes and produce higher resolution outputs. We also find that using viewer-centered coordinates is advantageous for novel objects, while object-centered representations are better for more familiar objects. Interestingly, the coordinate frame significantly affects the shape representation learned, with object-centered placing more importance on implicitly recognizing the object category and viewer-centered producing shape representations with less dependence on category recognition.

Tanmay Gupta, Dustin Schwenk, Ali Farhadi et al.

Imagining a scene described in natural language with realistic layout and appearance of entities is the ultimate test of spatial, visual, and semantic world knowledge. Towards this goal, we present the Composition, Retrieval, and Fusion Network (CRAFT), a model capable of learning this knowledge from video-caption data and applying it while generating videos from novel captions. CRAFT explicitly predicts a temporal-layout of mentioned entities (characters and objects), retrieves spatio-temporal entity segments from a video database and fuses them to generate scene videos. Our contributions include sequential training of components of CRAFT while jointly modeling layout and appearances, and losses that encourage learning compositional representations for retrieval. We evaluate CRAFT on semantic fidelity to caption, composition consistency, and visual quality. CRAFT outperforms direct pixel generation approaches and generalizes well to unseen captions and to unseen video databases with no text annotations. We demonstrate CRAFT on FLINTSTONES, a new richly annotated video-caption dataset with over 25000 videos. For a glimpse of videos generated by CRAFT, see https://youtu.be/688Vv86n0z8.

Chuhang Zou, Alex Colburn, Qi Shan et al.

We propose an algorithm to predict room layout from a single image that generalizes across panoramas and perspective images, cuboid layouts and more general layouts (e.g. L-shape room). Our method operates directly on the panoramic image, rather than decomposing into perspective images as do recent works. Our network architecture is similar to that of RoomNet, but we show improvements due to aligning the image based on vanishing points, predicting multiple layout elements (corners, boundaries, size and translation), and fitting a constrained Manhattan layout to the resulting predictions. Our method compares well in speed and accuracy to other existing work on panoramas, achieves among the best accuracy for perspective images, and can handle both cuboid-shaped and more general Manhattan layouts.

Chuhang Zou, Ruiqi Guo, Zhizhong Li et al.

One major goal of vision is to infer physical models of objects, surfaces, and their layout from sensors. In this paper, we aim to interpret indoor scenes from one RGBD image. Our representation encodes the layout of orthogonal walls and the extent of objects, modeled with CAD-like 3D shapes. We parse both the visible and occluded portions of the scene and all observable objects, producing a complete 3D parse. Such a scene interpretation is useful for robotics and visual reasoning, but difficult to produce due to the well-known challenge of segmentation, the high degree of occlusion, and the diversity of objects in indoor scenes. We take a data-driven approach, generating sets of potential object regions, matching to regions in training images, and transferring and aligning associated 3D models while encouraging fit to observations and spatial consistency. We use support inference to aid interpretation and propose a retrieval scheme that uses convolutional neural networks (CNNs) to classify regions and retrieve objects with similar shapes. We demonstrate the performance of our method on our newly annotated NYUd v2 dataset with detailed 3D shapes.

Joseph DeGol, Timothy Bretl, Derek Hoiem

Current fiducial marker detection algorithms rely on marker IDs for false positive rejection. Time is wasted on potential detections that will eventually be rejected as false positives. We introduce ChromaTag, a fiducial marker and detection algorithm designed to use opponent colors to limit and quickly reject initial false detections and grayscale for precise localization. Through experiments, we show that ChromaTag is significantly faster than current fiducial markers while achieving similar or better detection accuracy. We also show how tag size and viewing direction effect detection accuracy. Our contribution is significant because fiducial markers are often used in real-time applications (e.g. marker assisted robot navigation) where heavy computation is required by other parts of the system.

Chuhang Zou, Ersin Yumer, Jimei Yang et al.

The success of various applications including robotics, digital content creation, and visualization demand a structured and abstract representation of the 3D world from limited sensor data. Inspired by the nature of human perception of 3D shapes as a collection of simple parts, we explore such an abstract shape representation based on primitives. Given a single depth image of an object, we present 3D-PRNN, a generative recurrent neural network that synthesizes multiple plausible shapes composed of a set of primitives. Our generative model encodes symmetry characteristics of common man-made objects, preserves long-range structural coherence, and describes objects of varying complexity with a compact representation. We also propose a method based on Gaussian Fields to generate a large scale dataset of primitive-based shape representations to train our network. We evaluate our approach on a wide range of examples and show that it outperforms nearest-neighbor based shape retrieval methods and is on-par with voxel-based generative models while using a significantly reduced parameter space.

Tanmay Gupta, Kevin Shih, Saurabh Singh et al.

An important goal of computer vision is to build systems that learn visual representations over time that can be applied to many tasks. In this paper, we investigate a vision-language embedding as a core representation and show that it leads to better cross-task transfer than standard multi-task learning. In particular, the task of visual recognition is aligned to the task of visual question answering by forcing each to use the same word-region embeddings. We show this leads to greater inductive transfer from recognition to VQA than standard multitask learning. Visual recognition also improves, especially for categories that have relatively few recognition training labels but appear often in the VQA setting. Thus, our paper takes a small step towards creating more general vision systems by showing the benefit of interpretable, flexible, and trainable core representations.

Joseph DeGol, Mani Golparvar-Fard, Derek Hoiem

Our goal is to recognize material categories using images and geometry information. In many applications, such as construction management, coarse geometry information is available. We investigate how 3D geometry (surface normals, camera intrinsic and extrinsic parameters) can be used with 2D features (texture and color) to improve material classification. We introduce a new dataset, GeoMat, which is the first to provide both image and geometry data in the form of: (i) training and testing patches that were extracted at different scales and perspectives from real world examples of each material category, and (ii) a large scale construction site scene that includes 160 images and over 800,000 hand labeled 3D points. Our results show that using 2D and 3D features both jointly and independently to model materials improves classification accuracy across multiple scales and viewing directions for both material patches and images of a large scale construction site scene.

Zhizhong Li, Derek Hoiem

When building a unified vision system or gradually adding new capabilities to a system, the usual assumption is that training data for all tasks is always available. However, as the number of tasks grows, storing and retraining on such data becomes infeasible. A new problem arises where we add new capabilities to a Convolutional Neural Network (CNN), but the training data for its existing capabilities are unavailable. We propose our Learning without Forgetting method, which uses only new task data to train the network while preserving the original capabilities. Our method performs favorably compared to commonly used feature extraction and fine-tuning adaption techniques and performs similarly to multitask learning that uses original task data we assume unavailable. A more surprising observation is that Learning without Forgetting may be able to replace fine-tuning with similar old and new task datasets for improved new task performance.