Narendra Ahuja

Moitreya Chatterjee, Narendra Ahuja, Anoop Cherian

There exists an unequivocal distinction between the sound produced by a static source and that produced by a moving one, especially when the source moves towards or away from the microphone. In this paper, we propose to use this connection between audio and visual dynamics for solving two challenging tasks simultaneously, namely: (i) separating audio sources from a mixture using visual cues, and (ii) predicting the 3D visual motion of a sounding source using its separated audio. Towards this end, we present Audio Separator and Motion Predictor (ASMP) -- a deep learning framework that leverages the 3D structure of the scene and the motion of sound sources for better audio source separation. At the heart of ASMP is a 2.5D scene graph capturing various objects in the video and their pseudo-3D spatial proximities. This graph is constructed by registering together 2.5D monocular depth predictions from the 2D video frames and associating the 2.5D scene regions with the outputs of an object detector applied on those frames. The ASMP task is then mathematically modeled as the joint problem of: (i) recursively segmenting the 2.5D scene graph into several sub-graphs, each associated with a constituent sound in the input audio mixture (which is then separated) and (ii) predicting the 3D motions of the corresponding sound sources from the separated audio. To empirically evaluate ASMP, we present experiments on two challenging audio-visual datasets, viz. Audio Separation in the Wild (ASIW) and Audio Visual Event (AVE). Our results demonstrate that ASMP achieves a clear improvement in source separation quality, outperforming prior works on both datasets, while also estimating the direction of motion of the sound sources better than other methods.

Shubhang Bhatnagar, Sharath Gopal, Narendra Ahuja et al.

Gestures form an important medium of communication between humans and machines. An overwhelming majority of existing gesture recognition methods are tailored to a scenario where humans and machines are located very close to each other. This short-distance assumption does not hold true for several types of interactions, for example gesture-based interactions with a floor cleaning robot or with a drone. Methods made for short-distance recognition are unable to perform well on long-distance recognition due to gestures occupying only a small portion of the input data. Their performance is especially worse in resource constrained settings where they are not able to effectively focus their limited compute on the gesturing subject. We propose a novel, accurate and efficient method for the recognition of gestures from longer distances. It uses a dynamic neural network to select features from gesture-containing spatial regions of the input sensor data for further processing. This helps the network focus on features important for gesture recognition while discarding background features early on, thus making it more compute efficient compared to other techniques. We demonstrate the performance of our method on the LD-ConGR long-distance dataset where it outperforms previous state-of-the-art methods on recognition accuracy and compute efficiency.

Samyak Rawlekar, Amitabh Swain, Yujun Cai et al.

Self-supervised Vision Transformers (ViTs) like DINO show an emergent ability to discover objects, typically observed in [CLS] token attention maps of the final layer. However, these maps often contain spurious activations resulting in poor localization of objects. This is because the [CLS] token, trained on an image-level objective, summarizes the entire image instead of focusing on objects. This aggregation dilutes the object-centric information existing in the local, patch-level interactions. We analyze this by computing inter-patch similarity using patch-level attention components (query, key, and value) across all layers. We find that: (1) Object-centric properties are encoded in the similarity maps derived from all three components ($q, k, v$), unlike prior work that uses only key features or the [CLS] token. (2) This object-centric information is distributed across the network, not just confined to the final layer. Based on these insights, we introduce Object-DINO, a training-free method that extracts this distributed object-centric information. Object-DINO clusters attention heads across all layers based on the similarities of their patches and automatically identifies the object-centric cluster corresponding to all objects. We demonstrate Object-DINO's effectiveness on two applications: enhancing unsupervised object discovery (+3.6 to +12.4 CorLoc gains) and mitigating object hallucination in Multimodal Large Language Models by providing visual grounding. Our results demonstrate that using this distributed object-centric information improves downstream tasks without additional training.

Hao Zhang, Fang Li, Narendra Ahuja

In this paper, we address the challenge of decomposing Neural Radiance Fields (NeRF) into objects from an open vocabulary, a critical task for object manipulation in 3D reconstruction and view synthesis. Current techniques for NeRF decomposition involve a trade-off between the flexibility of processing open-vocabulary queries and the accuracy of 3D segmentation. We present, Open-vocabulary Embedded Neural Radiance Fields (Open-NeRF), that leverage large-scale, off-the-shelf, segmentation models like the Segment Anything Model (SAM) and introduce an integrate-and-distill paradigm with hierarchical embeddings to achieve both the flexibility of open-vocabulary querying and 3D segmentation accuracy. Open-NeRF first utilizes large-scale foundation models to generate hierarchical 2D mask proposals from varying viewpoints. These proposals are then aligned via tracking approaches and integrated within the 3D space and subsequently distilled into the 3D field. This process ensures consistent recognition and granularity of objects from different viewpoints, even in challenging scenarios involving occlusion and indistinct features. Our experimental results show that the proposed Open-NeRF outperforms state-of-the-art methods such as LERF \cite{lerf} and FFD \cite{ffd} in open-vocabulary scenarios. Open-NeRF offers a promising solution to NeRF decomposition, guided by open-vocabulary queries, enabling novel applications in robotics and vision-language interaction in open-world 3D scenes.

Shubham Kumar, Narendra Ahuja

Safety trained large language models (LLMs) can often be induced to answer harmful requests through jailbreak prompts. Because we lack a robust understanding of why LLMs are susceptible to jailbreaks, future frontier models operating more autonomously in higher-stakes settings may similarly be vulnerable to such attacks. Prior work has studied jailbreak success by examining the model's intermediate representations, identifying directions in this space that causally encode concepts like harmfulness and refusal. Then, they globally explain all jailbreak attacks as attempting to reduce or strengthen these concepts (e.g., reduce harmfulness). However, different jailbreak strategies may succeed by strengthening or suppressing different intermediate concepts, and the same jailbreak strategy may not work for different harmful request categories (e.g., violence vs. cyberattack); thus, we seek to give a local explanation -- i.e., why did this specific jailbreak succeed? To address this gap, we introduce LOCA, a method that gives Local, CAusal explanations of jailbreak success by identifying a minimal set of interpretable, intermediate representation changes that causally induce model refusal on an otherwise successful jailbreak request. We evaluate LOCA on harmful original-jailbreak pairs from a large jailbreak benchmark across Gemma and Llama chat models, comparing against prior methods adapted to this setting. LOCA can successfully induce refusal by making, on average, six interpretable changes; prior work routinely fails to achieve refusal even after 20 changes. LOCA is a step toward mechanistic, local explanations of jailbreak success in LLMs. Code to be released.

Samyak Rawlekar, Shubhang Bhatnagar, Narendra Ahuja

Vision-language models (VLMs) like CLIP have been adapted for Multi-Label Recognition (MLR) with partial annotations by leveraging prompt-learning, where positive and negative prompts are learned for each class to associate their embeddings with class presence or absence in the shared vision-text feature space. While this approach improves MLR performance by relying on VLM priors, we hypothesize that learning negative prompts may be suboptimal, as the datasets used to train VLMs lack image-caption pairs explicitly focusing on class absence. To analyze the impact of positive and negative prompt learning on MLR, we introduce PositiveCoOp and NegativeCoOp, where only one prompt is learned with VLM guidance while the other is replaced by an embedding vector learned directly in the shared feature space without relying on the text encoder. Through empirical analysis, we observe that negative prompts degrade MLR performance, and learning only positive prompts, combined with learned negative embeddings (PositiveCoOp), outperforms dual prompt learning approaches. Moreover, we quantify the performance benefits that prompt-learning offers over a simple vision-features-only baseline, observing that the baseline displays strong performance comparable to dual prompt learning approach (DualCoOp), when the proportion of missing labels is low, while requiring half the training compute and 16 times fewer parameters

Fang Li, Hao Zhang, Narendra Ahuja

Gaussian Splatting (GS) has significantly elevated scene reconstruction efficiency and novel view synthesis (NVS) accuracy compared to Neural Radiance Fields (NeRF), particularly for dynamic scenes. However, current 4D NVS methods, whether based on GS or NeRF, primarily rely on camera parameters provided by COLMAP and even utilize sparse point clouds generated by COLMAP for initialization, which lack accuracy as well are time-consuming. This sometimes results in poor dynamic scene representation, especially in scenes with large object movements, or extreme camera conditions e.g. small translations combined with large rotations. Some studies simultaneously optimize the estimation of camera parameters and scenes, supervised by additional information like depth, optical flow, etc. obtained from off-the-shelf models. Using this unverified information as ground truth can reduce robustness and accuracy, which does frequently occur for long monocular videos (with e.g. > hundreds of frames). We propose a novel approach that learns a high-fidelity 4D GS scene representation with self-calibration of camera parameters. It includes the extraction of 2D point features that robustly represent 3D structure, and their use for subsequent joint optimization of camera parameters and 3D structure towards overall 4D scene optimization. We demonstrate the accuracy and time efficiency of our method through extensive quantitative and qualitative experimental results on several standard benchmarks. The results show significant improvements over state-of-the-art methods for 4D novel view synthesis. The source code will be released soon at https://github.com/fangli333/SC-4DGS.

Aniket Shirke, Aziz Saifuddin, Achleshwar Luthra et al.

Increasing demand for meat products combined with farm labor shortages has resulted in a need to develop new real-time solutions to monitor animals effectively. Significant progress has been made in continuously locating individual pigs using tracking-by-detection methods. However, these methods fail for oblong pens because a single fixed camera does not cover the entire floor at adequate resolution. We address this problem by using multiple cameras, placed such that the visual fields of adjacent cameras overlap, and together they span the entire floor. Avoiding breaks in tracking requires inter-camera handover when a pig crosses from one camera's view into that of an adjacent camera. We identify the adjacent camera and the shared pig location on the floor at the handover time using inter-view homography. Our experiments involve two grow-finish pens, housing 16-17 pigs each, and three RGB cameras. Our algorithm first detects pigs using a deep learning-based object detection model (YOLO) and creates their local tracking IDs using a multi-object tracking algorithm (DeepSORT). We then use inter-camera shared locations to match multiple views and generate a global ID for each pig that holds throughout tracking. To evaluate our approach, we provide five two-minutes long video sequences with fully annotated global identities. We track pigs in a single camera view with a Multi-Object Tracking Accuracy and Precision of 65.0% and 54.3% respectively and achieve a Camera Handover Accuracy of 74.0%. We open-source our code and annotated dataset at https://github.com/AIFARMS/multi-camera-pig-tracking

Hao Zhang, Di Chang, Fang Li et al.

With the success of 2D and 3D visual generative models, there is growing interest in generating 4D content. Existing methods primarily rely on text prompts to produce 4D content, but they often fall short of accurately defining complex or rare motions. To address this limitation, we propose MagicPose4D, a novel framework for refined control over both appearance and motion in 4D generation. Unlike current 4D generation methods, MagicPose4D accepts monocular videos or mesh sequences as motion prompts, enabling precise and customizable motion control. MagicPose4D comprises two key modules: (i) Dual-Phase 4D Reconstruction Module, which operates in two phases. The first phase focuses on capturing the model's shape using accurate 2D supervision and less accurate but geometrically informative 3D pseudo-supervision without imposing skeleton constraints. The second phase extracts the 3D motion (skeleton poses) using more accurate pseudo-3D supervision, obtained in the first phase and introduces kinematic chain-based skeleton constraints to ensure physical plausibility. Additionally, we propose a Global-local Chamfer loss that aligns the overall distribution of predicted mesh vertices with the supervision while maintaining part-level alignment without extra annotations. (ii) Cross-category Motion Transfer Module, which leverages the extracted motion from the 4D reconstruction module and uses a kinematic-chain-based skeleton to achieve cross-category motion transfer. It ensures smooth transitions between frames through dynamic rigidity, facilitating robust generalization without additional training. Through extensive experiments, we demonstrate that MagicPose4D significantly improves the accuracy and consistency of 4D content generation, outperforming existing methods in various benchmarks.

Hao Zhang, Fang Li, Lu Qi et al.

Addressing Out-Of-Distribution (OOD) Segmentation and Zero-Shot Semantic Segmentation (ZS3) is challenging, necessitating segmenting unseen classes. Existing strategies adapt the class-agnostic Mask2Former (CA-M2F) tailored to specific tasks. However, these methods cater to singular tasks, demand training from scratch, and we demonstrate certain deficiencies in CA-M2F, which affect performance. We propose the Class-Agnostic Structure-Constrained Learning (CSL), a plug-in framework that can integrate with existing methods, thereby embedding structural constraints and achieving performance gain, including the unseen, specifically OOD, ZS3, and domain adaptation (DA) tasks. There are two schemes for CSL to integrate with existing methods (1) by distilling knowledge from a base teacher network, enforcing constraints across training and inference phrases, or (2) by leveraging established models to obtain per-pixel distributions without retraining, appending constraints during the inference phase. We propose soft assignment and mask split methodologies that enhance OOD object segmentation. Empirical evaluations demonstrate CSL's prowess in boosting the performance of existing algorithms spanning OOD segmentation, ZS3, and DA segmentation, consistently transcending the state-of-art across all three tasks.

Hao Zhang, Fang Li, Samyak Rawlekar et al.

Reconstructing dynamic articulated objects from a singular monocular video is challenging, requiring joint estimation of shape, motion, and camera parameters from limited views. Current methods typically demand extensive computational resources and training time, and require additional human annotations such as predefined parametric models, camera poses, and key points, limiting their generalizability. We propose Synergistic Shape and Skeleton Optimization (S3O), a novel two-phase method that forgoes these prerequisites and efficiently learns parametric models including visible shapes and underlying skeletons. Conventional strategies typically learn all parameters simultaneously, leading to interdependencies where a single incorrect prediction can result in significant errors. In contrast, S3O adopts a phased approach: it first focuses on learning coarse parametric models, then progresses to motion learning and detail addition. This method substantially lowers computational complexity and enhances robustness in reconstruction from limited viewpoints, all without requiring additional annotations. To address the current inadequacies in 3D reconstruction from monocular video benchmarks, we collected the PlanetZoo dataset. Our experimental evaluations on standard benchmarks and the PlanetZoo dataset affirm that S3O provides more accurate 3D reconstruction, and plausible skeletons, and reduces the training time by approximately 60% compared to the state-of-the-art, thus advancing the state of the art in dynamic object reconstruction.

Samyak Rawlekar, Shubhang Bhatnagar, Vishnuvardhan Pogunulu Srinivasulu et al.

Multi-label Recognition (MLR) involves the identification of multiple objects within an image. To address the additional complexity of this problem, recent works have leveraged information from vision-language models (VLMs) trained on large text-images datasets for the task. These methods learn an independent classifier for each object (class), overlooking correlations in their occurrences. Such co-occurrences can be captured from the training data as conditional probabilities between a pair of classes. We propose a framework to extend the independent classifiers by incorporating the co-occurrence information for object pairs to improve the performance of independent classifiers. We use a Graph Convolutional Network (GCN) to enforce the conditional probabilities between classes, by refining the initial estimates derived from image and text sources obtained using VLMs. We validate our method on four MLR datasets, where our approach outperforms all state-of-the-art methods.

Shubhang Bhatnagar, Narendra Ahuja

Unsupervised deep metric learning (UDML) focuses on learning a semantic representation space using only unlabeled data. This challenging problem requires accurately estimating the similarity between data points, which is used to supervise a deep network. For this purpose, we propose to model the high-dimensional data manifold using a piecewise-linear approximation, with each low-dimensional linear piece approximating the data manifold in a small neighborhood of a point. These neighborhoods are used to estimate similarity between data points. We empirically show that this similarity estimate correlates better with the ground truth than the similarity estimates of current state-of-the-art techniques. We also show that proxies, commonly used in supervised metric learning, can be used to model the piecewise-linear manifold in an unsupervised setting, helping improve performance. Our method outperforms existing unsupervised metric learning approaches on standard zero-shot image retrieval benchmarks.

Hao Zhang, Chun-Han Yao, Simon Donné et al.

We present Stable Part Diffusion 4D (SP4D), a framework for generating paired RGB and kinematic part videos from monocular inputs. Unlike conventional part segmentation methods that rely on appearance-based semantic cues, SP4D learns to produce kinematic parts - structural components aligned with object articulation and consistent across views and time. SP4D adopts a dual-branch diffusion model that jointly synthesizes RGB frames and corresponding part segmentation maps. To simplify the architecture and flexibly enable different part counts, we introduce a spatial color encoding scheme that maps part masks to continuous RGB-like images. This encoding allows the segmentation branch to share the latent VAE from the RGB branch, while enabling part segmentation to be recovered via straightforward post-processing. A Bidirectional Diffusion Fusion (BiDiFuse) module enhances cross-branch consistency, supported by a contrastive part consistency loss to promote spatial and temporal alignment of part predictions. We demonstrate that the generated 2D part maps can be lifted to 3D to derive skeletal structures and harmonic skinning weights with few manual adjustments. To train and evaluate SP4D, we construct KinematicParts20K, a curated dataset of over 20K rigged objects selected and processed from Objaverse XL (Deitke et al., 2023), each paired with multi-view RGB and part video sequences. Experiments show that SP4D generalizes strongly to diverse scenarios, including real-world videos, novel generated objects, and rare articulated poses, producing kinematic-aware outputs suitable for downstream animation and motion-related tasks.

Hao Zhang, Haolan Xu, Chun Feng et al.

Skinning and rigging are fundamental components in animation, articulated object reconstruction, motion transfer, and 4D generation. Existing approaches predominantly rely on Linear Blend Skinning (LBS), due to its simplicity and differentiability. However, LBS introduces artifacts such as volume loss and unnatural deformations, and it fails to model elastic materials like soft tissues, fur, and flexible appendages (e.g., elephant trunks, ears, and fatty tissues). In this work, we propose PhysRig: a differentiable physics-based skinning and rigging framework that overcomes these limitations by embedding the rigid skeleton into a volumetric representation (e.g., a tetrahedral mesh), which is simulated as a deformable soft-body structure driven by the animated skeleton. Our method leverages continuum mechanics and discretizes the object as particles embedded in an Eulerian background grid to ensure differentiability with respect to both material properties and skeletal motion. Additionally, we introduce material prototypes, significantly reducing the learning space while maintaining high expressiveness. To evaluate our framework, we construct a comprehensive synthetic dataset using meshes from Objaverse, The Amazing Animals Zoo, and MixaMo, covering diverse object categories and motion patterns. Our method consistently outperforms traditional LBS-based approaches, generating more realistic and physically plausible results. Furthermore, we demonstrate the applicability of our framework in the pose transfer task highlighting its versatility for articulated object modeling.

Sandeep Nagar, Mark Hasegawa-Johnson, David G. Beiser et al.

The COVID-19 pandemic has underscored the need for low-cost, scalable approaches to measuring contactless vital signs, either during initial triage at a healthcare facility or virtual telemedicine visits. Remote photoplethysmography (rPPG) can accurately estimate heart rate (HR) when applied to close-up videos of healthy volunteers in well-lit laboratory settings. However, results from such highly optimized laboratory studies may not be readily translated to healthcare settings. One significant barrier to the practical application of rPPG in health care is the accurate localization of the region of interest (ROI). Clinical or telemedicine visits may involve sub-optimal lighting, movement artifacts, variable camera angle, and subject distance. This paper presents an rPPG ROI selection method based on 3D facial landmarks and patient head yaw angle. We then demonstrate the robustness of this ROI selection method when coupled to the Plane-Orthogonal-to-Skin (POS) rPPG method when applied to videos of patients presenting to an Emergency Department for respiratory complaints. Our results demonstrate the effectiveness of our proposed approach in improving the accuracy and robustness of rPPG in a challenging clinical environment.

Dwip Dalal, Utkarsh Mishra, Narendra Ahuja et al.

Leveraging multimodal large language models (MLLMs) to develop embodied agents offers significant promise for addressing complex real-world tasks. However, current evaluation benchmarks remain predominantly language-centric or heavily reliant on simulated environments, rarely probing the nuanced, knowledge-intensive reasoning essential for practical, real-world scenarios. To bridge this critical gap, we introduce the task of Sparsely Grounded Visual Navigation, explicitly designed to evaluate the sequential decision-making abilities of MLLMs in challenging, knowledge-intensive real-world environment. We operationalize this task with CityNav, a comprehensive benchmark encompassing four diverse global cities, specifically constructed to assess raw MLLM-driven agents in city navigation. Agents are required to rely solely on visual inputs and internal multimodal reasoning to sequentially navigate 50+ decision points without additional environmental annotations or specialized architectural modifications. Crucially, agents must autonomously achieve localization through interpreting city-specific cues and recognizing landmarks, perform spatial reasoning, and strategically plan and execute routes to their destinations. Through extensive evaluations, we demonstrate that current state-of-the-art MLLMs, reasoning techniques (e.g., GEPA, chain-of-thought, reflection) and competitive baseline PReP significantly underperform in this challenging setting. To address this, we propose Verbalization of Path(VoP), which explicitly grounds the agent's internal reasoning by probing city-scale cognitive maps (key landmarks and directions toward the destination) from the MLLM, substantially enhancing navigation success. Project Webpage: https://dwipddalal.github.io/AgentNav/

Shubhang Bhatnagar, Andy Xu, Kar-Han Tan et al.

Large Language Models (LLMs) with multimodal capabilities have revolutionized vision-language tasks, but their deployment often requires huge memory and computational resources. While post-training quantization (PTQ) has successfully compressed language models to as low as 1-bit precision without significant performance loss, its effectiveness for multimodal LLMs (MLLMs) remains relatively unexplored. In this paper, we present the first study on ultra-low bit (<4-bit) quantization for multimodal LLMs. Our analysis reveals that multimodal tokens and intermediate layer activations produced by them exhibit significantly higher statistical variance and entropy compared to text tokens, making them less tolerant to ultra-low bit quantization. However, the activation distributions of multimodal tokens varies significantly over different layers, with some layers having lower entropy activation distributions. We empirically show that such layers in these models can better tolerate ultra-low bit quantization. Building on these insights, we propose a novel strategy for MLLM quantization, LUQ: Layerwise Ultra-Low Bit Quantization, which selectively applies ultra-low bit quantization to layers that are more resilient to it. Additionally, we also show that using a mix of multimodal tokens (image and text) for PTQ boosts VQA performance in the ultra-low bit regime. We evaluate our method on LLaVA-1.5 and Qwen-2.5-VL across 9 popular VQA benchmarks. The resulting LUQ models use 40% and 31% less memory than their 4-bit counterparts, respectively, while exhibiting a performance degradation of less than 10% on the MME benchmark.

Fang Li, Hao Zhang, Narendra Ahuja

Although COLMAP has long remained the predominant method for camera parameter optimization in static scenes, it is constrained by its lengthy runtime and reliance on ground truth (GT) motion masks for application to dynamic scenes. Many efforts attempted to improve it by incorporating more priors as supervision such as GT focal length, motion masks, 3D point clouds, camera poses, and metric depth, which, however, are typically unavailable in casually captured RGB videos. In this paper, we propose a novel method for more accurate and efficient camera parameter optimization in dynamic scenes solely supervised by a single RGB video, dubbed ROS-Cam. Our method consists of three key components: (1) Patch-wise Tracking Filters, to establish robust and maximally sparse hinge-like relations across the RGB video. (2) Outlier-aware Joint Optimization, for efficient camera parameter optimization by adaptive down-weighting of moving outliers, without reliance on motion priors. (3) A Two-stage Optimization Strategy, to enhance stability and optimization speed by a trade-off between the Softplus limits and convex minima in losses. We visually and numerically evaluate our camera estimates. To further validate accuracy, we feed the camera estimates into a 4D reconstruction method and assess the resulting 3D scenes, and rendered 2D RGB and depth maps. We perform experiments on 4 real-world datasets (NeRF-DS, DAVIS, iPhone, and TUM-dynamics) and 1 synthetic dataset (MPI-Sintel), demonstrating that our method estimates camera parameters more efficiently and accurately with a single RGB video as the only supervision.

Xiaodan Hu, Chuhang Zou, Suchen Wang et al.

Recent video action recognition methods have shown excellent performance by adapting large-scale pre-trained language-image models to the video domain. However, language models contain rich common sense priors - the scene contexts that humans use to constitute an understanding of objects, human-object interactions, and activities - that have not been fully exploited. In this paper, we introduce a framework incorporating language-driven common sense priors to identify cluttered video action sequences from monocular views that are often heavily occluded. We propose: (1) A video context summary component that generates candidate objects, activities, and the interactions between objects and activities; (2) A description generation module that describes the current scene given the context and infers subsequent activities, through auxiliary prompts and common sense reasoning; (3) A multi-modal activity recognition head that combines visual and textual cues to recognize video actions. We demonstrate the effectiveness of our approach on the challenging Action Genome and Charades datasets.

Shubham Kumar, Narendra Ahuja

Deep vision models perform input-output computations that are hard to interpret. Concept-based explanation methods (CBEMs) increase interpretability by re-expressing parts of the model with human-understandable semantic units, or concepts. Checking if the derived explanations are faithful -- that is, they represent the model's internal computation -- requires a surrogate that combines concepts to compute the output. Simplifications made for interpretability inevitably reduce faithfulness, resulting in a tradeoff between the two. State-of-the-art unsupervised CBEMs (U-CBEMs) have reported increasingly interpretable concepts, while also being more faithful to the model. However, we observe that the reported improvement in faithfulness artificially results from either (1) using overly complex surrogates, which introduces an unmeasured cost to the explanation's interpretability, or (2) relying on deletion-based approaches that, as we demonstrate, do not properly measure faithfulness. We propose Surrogate Faithfulness (SURF), which (1) replaces prior complex surrogates with a simple, linear surrogate that measures faithfulness without changing the explanation's interpretability and (2) introduces well-motivated metrics that assess loss across all output classes, not just the predicted class. We validate SURF with a measure-over-measure study by proposing a simple sanity check -- explanations with random concepts should be less faithful -- which prior surrogates fail. SURF enables the first reliable faithfulness benchmark of U-CBEMs, revealing that many visually compelling U-CBEMs are not faithful. Code to be released.

Samyak Rawlekar, Yujun Cai, Yiwei Wang et al.

Vision-language models like CLIP excel at recognizing the single, prominent object in a scene. However, they struggle in complex scenes containing multiple objects. We identify a fundamental reason for this limitation: VLM feature space exhibits excessive mutual feature information (MFI), where the features of one class contain substantial information about other, unrelated classes. This high MFI becomes evident during class-specific queries, as unrelated objects are activated alongside the queried class. To address this limitation, we propose DCLIP, an efficient framework that learns an optimal level of mutual information while adding only minimal learnable parameters to a frozen VLM. DCLIP uses two complementary losses: a novel MFI Loss that regulates class feature similarity to prevent excessive overlap while preserving necessary shared information, and the Asymmetric Loss (ASL) that aligns image features with the disentangled text features. Through this disentanglement, DCLIP reduces excessive inter-class similarity by 30%. On multi-label recognition, DCLIP performs favorably over SOTA approaches on VOC2007 and COCO-14 while using 75% fewer training parameters. For zero-shot semantic segmentation, it shows improved performance across six benchmark datasets. These results highlight the importance of feature disentanglement for multi-object perception in VLMs.

Hao Zhang, Fang Li, Samyak Rawlekar et al.

3D Reconstruction of moving articulated objects without additional information about object structure is a challenging problem. Current methods overcome such challenges by employing category-specific skeletal models. Consequently, they do not generalize well to articulated objects in the wild. We treat an articulated object as an unknown, semi-rigid skeletal structure surrounded by nonrigid material (e.g., skin). Our method simultaneously estimates the visible (explicit) representation (3D shapes, colors, camera parameters) and the implicit skeletal representation, from motion cues in the object video without 3D supervision. Our implicit representation consists of four parts. (1) Skeleton, which specifies how semi-rigid parts are connected. (2) \textcolor{black}{Skinning Weights}, which associates each surface vertex with semi-rigid parts with probability. (3) Rigidity Coefficients, specifying the articulation of the local surface. (4) Time-Varying Transformations, which specify the skeletal motion and surface deformation parameters. We introduce an algorithm that uses physical constraints as regularization terms and iteratively estimates both implicit and explicit representations. Our method is category-agnostic, thus eliminating the need for category-specific skeletons, we show that our method outperforms state-of-the-art across standard video datasets.

Moitreya Chatterjee, Narendra Ahuja, Anoop Cherian

Predicting the future frames of a video is a challenging task, in part due to the underlying stochastic real-world phenomena. Prior approaches to solve this task typically estimate a latent prior characterizing this stochasticity, however do not account for the predictive uncertainty of the (deep learning) model. Such approaches often derive the training signal from the mean-squared error (MSE) between the generated frame and the ground truth, which can lead to sub-optimal training, especially when the predictive uncertainty is high. Towards this end, we introduce Neural Uncertainty Quantifier (NUQ) - a stochastic quantification of the model's predictive uncertainty, and use it to weigh the MSE loss. We propose a hierarchical, variational framework to derive NUQ in a principled manner using a deep, Bayesian graphical model. Our experiments on four benchmark stochastic video prediction datasets show that our proposed framework trains more effectively compared to the state-of-the-art models (especially when the training sets are small), while demonstrating better video generation quality and diversity against several evaluation metrics.

Moitreya Chatterjee, Jonathan Le Roux, Narendra Ahuja et al.

State-of-the-art approaches for visually-guided audio source separation typically assume sources that have characteristic sounds, such as musical instruments. These approaches often ignore the visual context of these sound sources or avoid modeling object interactions that may be useful to better characterize the sources, especially when the same object class may produce varied sounds from distinct interactions. To address this challenging problem, we propose Audio Visual Scene Graph Segmenter (AVSGS), a novel deep learning model that embeds the visual structure of the scene as a graph and segments this graph into subgraphs, each subgraph being associated with a unique sound obtained by co-segmenting the audio spectrogram. At its core, AVSGS uses a recursive neural network that emits mutually-orthogonal sub-graph embeddings of the visual graph using multi-head attention. These embeddings are used for conditioning an audio encoder-decoder towards source separation. Our pipeline is trained end-to-end via a self-supervised task consisting of separating audio sources using the visual graph from artificially mixed sounds. In this paper, we also introduce an "in the wild'' video dataset for sound source separation that contains multiple non-musical sources, which we call Audio Separation in the Wild (ASIW). This dataset is adapted from the AudioCaps dataset, and provides a challenging, natural, and daily-life setting for source separation. Thorough experiments on the proposed ASIW and the standard MUSIC datasets demonstrate state-of-the-art sound separation performance of our method against recent prior approaches.

Xiaodan Hu, Narendra Ahuja

Dance experts often view dance as a hierarchy of information, spanning low-level (raw images, image sequences), mid-levels (human poses and bodypart movements), and high-level (dance genre). We propose a Hierarchical Dance Video Recognition framework (HDVR). HDVR estimates 2D pose sequences, tracks dancers, and then simultaneously estimates corresponding 3D poses and 3D-to-2D imaging parameters, without requiring ground truth for 3D poses. Unlike most methods that work on a single person, our tracking works on multiple dancers, under occlusions. From the estimated 3D pose sequence, HDVR extracts body part movements, and therefrom dance genre. The resulting hierarchical dance representation is explainable to experts. To overcome noise and interframe correspondence ambiguities, we enforce spatial and temporal motion smoothness and photometric continuity over time. We use an LSTM network to extract 3D movement subsequences from which we recognize the dance genre. For experiments, we have identified 154 movement types, of 16 body parts, and assembled a new University of Illinois Dance (UID) Dataset, containing 1143 video clips of 9 genres covering 30 hours, annotated with movement and genre labels. Our experimental results demonstrate that our algorithms outperform the state-of-the-art 3D pose estimation methods, which also enhances our dance recognition performance.

Anoop Cherian, Moitreya Chatterjee, Narendra Ahuja

Learning associations across modalities is critical for robust multimodal reasoning, especially when a modality may be missing during inference. In this paper, we study this problem in the context of audio-conditioned visual synthesis -- a task that is important, for example, in occlusion reasoning. Specifically, our goal is to generate future video frames and their motion dynamics conditioned on audio and a few past frames. To tackle this problem, we present Sound2Sight, a deep variational framework, that is trained to learn a per frame stochastic prior conditioned on a joint embedding of audio and past frames. This embedding is learned via a multi-head attention-based audio-visual transformer encoder. The learned prior is then sampled to further condition a video forecasting module to generate future frames. The stochastic prior allows the model to sample multiple plausible futures that are consistent with the provided audio and the past context. Moreover, to improve the quality and coherence of the generated frames, we propose a multimodal discriminator that differentiates between a synthesized and a real audio-visual clip. We empirically evaluate our approach, vis-á-vis closely-related prior methods, on two new datasets viz. (i) Multimodal Stochastic Moving MNIST with a Surprise Obstacle, (ii) Youtube Paintings; as well as on the existing Audio-Set Drums dataset. Our extensive experiments demonstrate that Sound2Sight significantly outperforms the state of the art in the generated video quality, while also producing diverse video content.

Abhimanyu Dubey, Moitreya Chatterjee, Narendra Ahuja

We propose a novel Convolutional Neural Network (CNN) compression algorithm based on coreset representations of filters. We exploit the redundancies extant in the space of CNN weights and neuronal activations (across samples) in order to obtain compression. Our method requires no retraining, is easy to implement, and obtains state-of-the-art compression performance across a wide variety of CNN architectures. Coupled with quantization and Huffman coding, we create networks that provide AlexNet-like accuracy, with a memory footprint that is $832\times$ smaller than the original AlexNet, while also introducing significant reductions in inference time as well. Additionally these compressed networks when fine-tuned, successfully generalize to other domains as well.

Po-Han Huang, Kevin Matzen, Johannes Kopf et al.

We present DeepMVS, a deep convolutional neural network (ConvNet) for multi-view stereo reconstruction. Taking an arbitrary number of posed images as input, we first produce a set of plane-sweep volumes and use the proposed DeepMVS network to predict high-quality disparity maps. The key contributions that enable these results are (1) supervised pretraining on a photorealistic synthetic dataset, (2) an effective method for aggregating information across a set of unordered images, and (3) integrating multi-layer feature activations from the pre-trained VGG-19 network. We validate the efficacy of DeepMVS using the ETH3D Benchmark. Our results show that DeepMVS compares favorably against state-of-the-art conventional MVS algorithms and other ConvNet based methods, particularly for near-textureless regions and thin structures.

Yijun Li, Jia-Bin Huang, Narendra Ahuja et al.

Joint image filters leverage the guidance image as a prior and transfer the structural details from the guidance image to the target image for suppressing noise or enhancing spatial resolution. Existing methods either rely on various explicit filter constructions or hand-designed objective functions, thereby making it difficult to understand, improve, and accelerate these filters in a coherent framework. In this paper, we propose a learning-based approach for constructing joint filters based on Convolutional Neural Networks. In contrast to existing methods that consider only the guidance image, the proposed algorithm can selectively transfer salient structures that are consistent with both guidance and target images. We show that the model trained on a certain type of data, e.g., RGB and depth images, generalizes well to other modalities, e.g., flash/non-Flash and RGB/NIR images. We validate the effectiveness of the proposed joint filter through extensive experimental evaluations with state-of-the-art methods.

Shun Zhang, Jia-Bin Huang, Jongwoo Lim et al.

Multi-face tracking in unconstrained videos is a challenging problem as faces of one person often appear drastically different in multiple shots due to significant variations in scale, pose, expression, illumination, and make-up. Existing multi-target tracking methods often use low-level features which are not sufficiently discriminative for identifying faces with such large appearance variations. In this paper, we tackle this problem by learning discriminative, video-specific face representations using convolutional neural networks (CNNs). Unlike existing CNN-based approaches which are only trained on large-scale face image datasets offline, we use the contextual constraints to generate a large number of training samples for a given video, and further adapt the pre-trained face CNN to specific videos using discovered training samples. Using these training samples, we optimize the embedding space so that the Euclidean distances correspond to a measure of semantic face similarity via minimizing a triplet loss function. With the learned discriminative features, we apply the hierarchical clustering algorithm to link tracklets across multiple shots to generate trajectories. We extensively evaluate the proposed algorithm on two sets of TV sitcoms and YouTube music videos, analyze the contribution of each component, and demonstrate significant performance improvement over existing techniques.

Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja et al.

Convolutional neural networks have recently demonstrated high-quality reconstruction for single image super-resolution. However, existing methods often require a large number of network parameters and entail heavy computational loads at runtime for generating high-accuracy super-resolution results. In this paper, we propose the deep Laplacian Pyramid Super-Resolution Network for fast and accurate image super-resolution. The proposed network progressively reconstructs the sub-band residuals of high-resolution images at multiple pyramid levels. In contrast to existing methods that involve the bicubic interpolation for pre-processing (which results in large feature maps), the proposed method directly extracts features from the low-resolution input space and thereby entails low computational loads. We train the proposed network with deep supervision using the robust Charbonnier loss functions and achieve high-quality image reconstruction. Furthermore, we utilize the recursive layers to share parameters across as well as within pyramid levels, and thus drastically reduce the number of parameters. Extensive quantitative and qualitative evaluations on benchmark datasets show that the proposed algorithm performs favorably against the state-of-the-art methods in terms of run-time and image quality.

Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja et al.

Convolutional neural networks have recently demonstrated high-quality reconstruction for single-image super-resolution. In this paper, we propose the Laplacian Pyramid Super-Resolution Network (LapSRN) to progressively reconstruct the sub-band residuals of high-resolution images. At each pyramid level, our model takes coarse-resolution feature maps as input, predicts the high-frequency residuals, and uses transposed convolutions for upsampling to the finer level. Our method does not require the bicubic interpolation as the pre-processing step and thus dramatically reduces the computational complexity. We train the proposed LapSRN with deep supervision using a robust Charbonnier loss function and achieve high-quality reconstruction. Furthermore, our network generates multi-scale predictions in one feed-forward pass through the progressive reconstruction, thereby facilitates resource-aware applications. Extensive quantitative and qualitative evaluations on benchmark datasets show that the proposed algorithm performs favorably against the state-of-the-art methods in terms of speed and accuracy.

Xing Wei, Qingxiong Yang, Yihong Gong et al.

Superpixel segmentation is becoming ubiquitous in computer vision. In practice, an object can either be represented by a number of segments in finer levels of detail or included in a surrounding region at coarser levels of detail, and thus a superpixel segmentation hierarchy is useful for applications that require different levels of image segmentation detail depending on the particular image objects segmented. Unfortunately, there is no method that can generate all scales of superpixels accurately in real-time. As a result, a simple yet effective algorithm named Super Hierarchy (SH) is proposed in this paper. It is as accurate as the state-of-the-art but 1-2 orders of magnitude faster. The proposed method can be directly integrated with recent efficient edge detectors like the structured forest edges to significantly outperforms the state-of-the-art in terms of segmentation accuracy. Quantitative and qualitative evaluation on a number of computer vision applications was conducted, demonstrating that the proposed method is the top performer.