Makarand Tapaswi

Pierre-Louis Guhur, Shizhe Chen, Ricardo Garcia et al.

In human environments, robots are expected to accomplish a variety of manipulation tasks given simple natural language instructions. Yet, robotic manipulation is extremely challenging as it requires fine-grained motor control, long-term memory as well as generalization to previously unseen tasks and environments. To address these challenges, we propose a unified transformer-based approach that takes into account multiple inputs. In particular, our transformer architecture integrates (i) natural language instructions and (ii) multi-view scene observations while (iii) keeping track of the full history of observations and actions. Such an approach enables learning dependencies between history and instructions and improves manipulation precision using multiple views. We evaluate our method on the challenging RLBench benchmark and on a real-world robot. Notably, our approach scales to 74 diverse RLBench tasks and outperforms the state of the art. We also address instruction-conditioned tasks and demonstrate excellent generalization to previously unseen variations.

Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi et al.

Localizing objects in 3D scenes based on natural language requires understanding and reasoning about spatial relations. In particular, it is often crucial to distinguish similar objects referred by the text, such as "the left most chair" and "a chair next to the window". In this work we propose a language-conditioned transformer model for grounding 3D objects and their spatial relations. To this end, we design a spatial self-attention layer that accounts for relative distances and orientations between objects in input 3D point clouds. Training such a layer with visual and language inputs enables to disambiguate spatial relations and to localize objects referred by the text. To facilitate the cross-modal learning of relations, we further propose a teacher-student approach where the teacher model is first trained using ground-truth object labels, and then helps to train a student model using point cloud inputs. We perform ablation studies showing advantages of our approach. We also demonstrate our model to significantly outperform the state of the art on the challenging Nr3D, Sr3D and ScanRefer 3D object grounding datasets.

Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi et al.

In vision-and-language navigation (VLN), an embodied agent is required to navigate in realistic 3D environments following natural language instructions. One major bottleneck for existing VLN approaches is the lack of sufficient training data, resulting in unsatisfactory generalization to unseen environments. While VLN data is typically collected manually, such an approach is expensive and prevents scalability. In this work, we address the data scarcity issue by proposing to automatically create a large-scale VLN dataset from 900 unlabeled 3D buildings from HM3D. We generate a navigation graph for each building and transfer object predictions from 2D to generate pseudo 3D object labels by cross-view consistency. We then fine-tune a pretrained language model using pseudo object labels as prompts to alleviate the cross-modal gap in instruction generation. Our resulting HM3D-AutoVLN dataset is an order of magnitude larger than existing VLN datasets in terms of navigation environments and instructions. We experimentally demonstrate that HM3D-AutoVLN significantly increases the generalization ability of resulting VLN models. On the SPL metric, our approach improves over state of the art by 7.1% and 8.1% on the unseen validation splits of REVERIE and SOON datasets respectively.

Piyush Bagad, Makarand Tapaswi, Cees G. M. Snoek

Modelling and understanding time remains a challenge in contemporary video understanding models. With language emerging as a key driver towards powerful generalization, it is imperative for foundational video-language models to have a sense of time. In this paper, we consider a specific aspect of temporal understanding: consistency of time order as elicited by before/after relations. We establish that seven existing video-language models struggle to understand even such simple temporal relations. We then question whether it is feasible to equip these foundational models with temporal awareness without re-training them from scratch. Towards this, we propose a temporal adaptation recipe on top of one such model, VideoCLIP, based on post-pretraining on a small amount of video-text data. We conduct a zero-shot evaluation of the adapted models on six datasets for three downstream tasks which require varying degrees of time awareness. We observe encouraging performance gains especially when the task needs higher time awareness. Our work serves as a first step towards probing and instilling a sense of time in existing video-language models without the need for data and compute-intense training from scratch.

Manu Gaur, Darshan Singh S, Makarand Tapaswi

Visual Question Answering (VQA) with multiple choice questions enables a vision-centric evaluation of Multimodal Large Language Models (MLLMs). Although it reliably checks the existence of specific visual abilities, it is easier for the model to select an answer from multiple choices (VQA evaluation) than to generate the answer itself. In this work, we offer a novel perspective: we evaluate how well an MLLM understands a specific visual concept by its ability to uniquely describe two extremely similar images that differ only in the targeted visual concept. Specifically, we assess the ability of MLLMs to capture specific points of visual differences using self-retrieval, i.e., by retrieving the target image using its generated caption against the other image in the pair serving as the distractor. We curate 247 highly similar image pairs as part of the D3 benchmark. For each image pair, the model is prompted to: (1) Detect a specific visual difference, and (2) Describe the target image uniquely such that it (3) Discriminates the target image from the distractor. Self-retrieval within D3 enables whitebox evaluation across six different visual patterns, revealing that current models struggle to independently discern fine-grained visual differences, with open-source models failing to outperform random guess.

Manu Gaur, Darshan Singh, Makarand Tapaswi

Image captioning systems are unable to generate fine-grained captions as they are trained on data that is either noisy (alt-text) or generic (human annotations). This is further exacerbated by maximum likelihood training that encourages generation of frequently occurring phrases. Previous works have tried to address this limitation by fine-tuning captioners with a self-retrieval (SR) reward. However, we find that SR fine-tuning has a tendency to reduce caption faithfulness and even hallucinate. In this work, we circumvent this bottleneck by improving the MLE initialization of the captioning system and designing a curriculum for the SR fine-tuning process. To this extent, we present (1) Visual Caption Boosting, a novel framework to instill fine-grainedness in generic image captioning datasets while remaining anchored in human annotations; and (2) BagCurri, a carefully designed training curriculum that more optimally leverages the contrastive nature of the self-retrieval reward. Jointly, they enable the captioner to describe fine-grained aspects in the image while preserving faithfulness to ground-truth captions. Our approach outperforms previous work by +8.9% on SR against 99 random distractors (RD100) (Dessi et al., 2023); and +7.6% on ImageCoDe. Additionally, existing metrics to evaluate captioning systems fail to reward diversity or evaluate a model's fine-grained understanding ability. Our third contribution addresses this by proposing self-retrieval from the lens of evaluation. We introduce TrueMatch, a benchmark comprising bags of highly similar images that uses SR to assess the captioner's ability to capture subtle visual distinctions. We evaluate and compare several state-of-the-art open-source MLLMs on TrueMatch, and find that our SR approach outperforms them all by a significant margin (e.g. +4.8% - 7.1% over Cambrian) while having 1-2 orders of magnitude fewer parameters.

Vladimir Petrik, Mohammad Nomaan Qureshi, Josef Sivic et al.

We aim to teach robots to perform simple object manipulation tasks by watching a single video demonstration. Towards this goal, we propose an optimization approach that outputs a coarse and temporally evolving 3D scene to mimic the action demonstrated in the input video. Similar to previous work, a differentiable renderer ensures perceptual fidelity between the 3D scene and the 2D video. Our key novelty lies in the inclusion of a differentiable approach to solve a set of Ordinary Differential Equations (ODEs) that allows us to approximately model laws of physics such as gravity, friction, and hand-object or object-object interactions. This not only enables us to dramatically improve the quality of estimated hand and object states, but also produces physically admissible trajectories that can be directly translated to a robot without the need for costly reinforcement learning. We evaluate our approach on a 3D reconstruction task that consists of 54 video demonstrations sourced from 9 actions such as pull something from right to left or put something in front of something. Our approach improves over previous state-of-the-art by almost 30%, demonstrating superior quality on especially challenging actions involving physical interactions of two objects such as put something onto something. Finally, we showcase the learned skills on a Franka Emika Panda robot.

Dhaval Taunk, Lakshya Khanna, Pavan Kandru et al.

Commonsense question-answering (QA) methods combine the power of pre-trained Language Models (LM) with the reasoning provided by Knowledge Graphs (KG). A typical approach collects nodes relevant to the QA pair from a KG to form a Working Graph (WG) followed by reasoning using Graph Neural Networks(GNNs). This faces two major challenges: (i) it is difficult to capture all the information from the QA in the WG, and (ii) the WG contains some irrelevant nodes from the KG. To address these, we propose GrapeQA with two simple improvements on the WG: (i) Prominent Entities for Graph Augmentation identifies relevant text chunks from the QA pair and augments the WG with corresponding latent representations from the LM, and (ii) Context-Aware Node Pruning removes nodes that are less relevant to the QA pair. We evaluate our results on OpenBookQA, CommonsenseQA and MedQA-USMLE and see that GrapeQA shows consistent improvements over its LM + KG predecessor (QA-GNN in particular) and large improvements on OpenBookQA.

Zeeshan Khan, C. V. Jawahar, Makarand Tapaswi

Dense video understanding requires answering several questions such as who is doing what to whom, with what, how, why, and where. Recently, Video Situation Recognition (VidSitu) is framed as a task for structured prediction of multiple events, their relationships, and actions and various verb-role pairs attached to descriptive entities. This task poses several challenges in identifying, disambiguating, and co-referencing entities across multiple verb-role pairs, but also faces some challenges of evaluation. In this work, we propose the addition of spatio-temporal grounding as an essential component of the structured prediction task in a weakly supervised setting, and present a novel three stage Transformer model, VideoWhisperer, that is empowered to make joint predictions. In stage one, we learn contextualised embeddings for video features in parallel with key objects that appear in the video clips to enable fine-grained spatio-temporal reasoning. The second stage sees verb-role queries attend and pool information from object embeddings, localising answers to questions posed about the action. The final stage generates these answers as captions to describe each verb-role pair present in the video. Our model operates on a group of events (clips) simultaneously and predicts verbs, verb-role pairs, their nouns, and their grounding on-the-fly. When evaluated on a grounding-augmented version of the VidSitu dataset, we observe a large improvement in entity captioning accuracy, as well as the ability to localize verb-roles without grounding annotations at training time.

Dhruv Srivastava, Aditya Kumar Singh, Makarand Tapaswi

Movie story analysis requires understanding characters' emotions and mental states. Towards this goal, we formulate emotion understanding as predicting a diverse and multi-label set of emotions at the level of a movie scene and for each character. We propose EmoTx, a multimodal Transformer-based architecture that ingests videos, multiple characters, and dialog utterances to make joint predictions. By leveraging annotations from the MovieGraphs dataset, we aim to predict classic emotions (e.g. happy, angry) and other mental states (e.g. honest, helpful). We conduct experiments on the most frequently occurring 10 and 25 labels, and a mapping that clusters 181 labels to 26. Ablation studies and comparison against adapted state-of-the-art emotion recognition approaches shows the effectiveness of EmoTx. Analyzing EmoTx's self-attention scores reveals that expressive emotions often look at character tokens while other mental states rely on video and dialog cues.

Prajneya Kumar, Eshika Khandelwal, Makarand Tapaswi et al.

Understanding what makes a video memorable has important applications in advertising or education technology. Towards this goal, we investigate spatio-temporal attention mechanisms underlying video memorability. Different from previous works that fuse multiple features, we adopt a simple CNN+Transformer architecture that enables analysis of spatio-temporal attention while matching state-of-the-art (SoTA) performance on video memorability prediction. We compare model attention against human gaze fixations collected through a small-scale eye-tracking study where humans perform the video memory task. We uncover the following insights: (i) Quantitative saliency metrics show that our model, trained only to predict a memorability score, exhibits similar spatial attention patterns to human gaze, especially for more memorable videos. (ii) The model assigns greater importance to initial frames in a video, mimicking human attention patterns. (iii) Panoptic segmentation reveals that both (model and humans) assign a greater share of attention to things and less attention to stuff as compared to their occurrence probability.

Arsh Verma, Makarand Tapaswi

Chest radiograph (or Chest X-Ray, CXR) is a popular medical imaging modality that is used by radiologists across the world to diagnose heart or lung conditions. Over the last decade, Convolutional Neural Networks (CNN), have seen success in identifying pathologies in CXR images. Typically, these CNNs are pretrained on the standard ImageNet classification task, but this assumes availability of large-scale annotated datasets. In this work, we analyze the utility of pretraining on unlabeled ImageNet or Chest X-Ray (CXR) datasets using various algorithms and in multiple settings. Some findings of our work include: (i) supervised training with labeled ImageNet learns strong representations that are hard to beat; (ii) self-supervised pretraining on ImageNet (~1M images) shows performance similar to self-supervised pretraining on a CXR dataset (~100K images); and (iii) the CNN trained on supervised ImageNet can be trained further with self-supervised CXR images leading to improvements, especially when the downstream dataset is on the order of a few thousand images.

Jaidev Shriram, Makarand Tapaswi, Vinoo Alluri

Reading, much like music listening, is an immersive experience that transports readers while taking them on an emotional journey. Listening to complementary music has the potential to amplify the reading experience, especially when the music is stylistically cohesive and emotionally relevant. In this paper, we propose the first fully automatic method to build a dense soundtrack for books, which can play high-quality instrumental music for the entirety of the reading duration. Our work employs a unique text processing and music weaving pipeline that determines the context and emotional composition of scenes in a chapter. This allows our method to identify and play relevant excerpts from the soundtrack of the book's movie adaptation. By relying on the movie composer's craftsmanship, our book soundtracks include expert-made motifs and other scene-specific musical characteristics. We validate the design decisions of our approach through a perceptual study. Our readers note that the book soundtrack greatly enhanced their reading experience, due to high immersiveness granted via uninterrupted and style-consistent music, and a heightened emotional state attained via high precision emotion and scene context recognition.

Aroof Aimen, Arsh Verma, Makarand Tapaswi et al.

Real-world application of chest X-ray abnormality classification requires dealing with several challenges: (i) limited training data; (ii) training and evaluation sets that are derived from different domains; and (iii) classes that appear during training may have partial overlap with classes of interest during evaluation. To address these challenges, we present an integrated framework called Generalized Cross-Domain Multi-Label Few-Shot Learning (GenCDML-FSL). The framework supports overlap in classes during training and evaluation, cross-domain transfer, adopts meta-learning to learn using few training samples, and assumes each chest X-ray image is either normal or associated with one or more abnormalities. Furthermore, we propose Generalized Episodic Training (GenET), a training strategy that equips models to operate with multiple challenges observed in the GenCDML-FSL scenario. Comparisons with well-established methods such as transfer learning, hybrid transfer learning, and multi-label meta-learning on multiple datasets show the superiority of our approach.

Abhinav Raundhal, Gaurav Behera, P J Narayanan et al.

Text as signs, labels, or instructions is a critical element of real-world scenes as they can convey important contextual information. 3D representations such as 3D Gaussian Splatting (3DGS) struggle to preserve fine-grained text details, while achieving high visual fidelity. Small errors in textual element reconstruction can lead to significant semantic loss. We propose STRinGS, a text-aware, selective refinement framework to address this issue for 3DGS reconstruction. Our method treats text and non-text regions separately, refining text regions first and merging them with non-text regions later for full-scene optimization. STRinGS produces sharp, readable text even in challenging configurations. We introduce a text readability measure OCR Character Error Rate (CER) to evaluate the efficacy on text regions. STRinGS results in a 63.6% relative improvement over 3DGS at just 7K iterations. We also introduce a curated dataset STRinGS-360 with diverse text scenarios to evaluate text readability in 3D reconstruction. Our method and dataset together push the boundaries of 3D scene understanding in text-rich environments, paving the way for more robust text-aware reconstruction methods.

Shivank Saxena, Dhruv Srivastava, Makarand Tapaswi

Recent advances in text-to-image models have enabled a new era of creative and controllable image generation. However, generating compositional scenes with multiple subjects and attributes remains a significant challenge. To enhance user control over subject placement, several layout-guided methods have been proposed. However, these methods face numerous challenges, particularly in compositional scenes. Unintended subjects often appear outside the layouts, generated images can be out-of-distribution and contain unnatural artifacts, or attributes bleed across subjects, leading to incorrect visual outputs. In this work, we propose MALeR, a method that addresses each of these challenges. Given a text prompt and corresponding layouts, our method prevents subjects from appearing outside the given layouts while being in-distribution. Additionally, we propose a masked, attribute-aware binding mechanism that prevents attribute leakage, enabling accurate rendering of subjects with multiple attributes, even in complex compositional scenes. Qualitative and quantitative evaluation demonstrates that our method achieves superior performance in compositional accuracy, generation consistency, and attribute binding compared to previous work. MALeR is particularly adept at generating images of scenes with multiple subjects and multiple attributes per subject.

Kawshik Manikantan, Makarand Tapaswi, Vineet Gandhi et al.

Recent evaluations of LLMs on coreference resolution have revealed that traditional output formats and evaluation metrics do not fully capture the models' referential understanding. To address this, we introduce IdentifyMe, a new benchmark for mention resolution presented in a multiple-choice question (MCQ) format, commonly used for evaluating LLMs. IdentifyMe features long narratives and employs heuristics to exclude easily identifiable mentions, creating a more challenging task. The benchmark also consists of a curated mixture of different mention types and corresponding entities, allowing for a fine-grained analysis of model performance. We evaluate both closed- and open source LLMs on IdentifyMe and observe a significant performance gap (20-30%) between the state-of-the-art sub-10B open models vs. closed ones. We observe that pronominal mentions, which have limited surface information, are typically much harder for models to resolve than nominal mentions. Additionally, we find that LLMs often confuse entities when their mentions overlap in nested structures. The highest-scoring model, GPT-4o, achieves 81.9% accuracy, highlighting the strong referential capabilities of state-of-the-art LLMs while also indicating room for further improvement.

Rahul Vigneswaran, Marc T. Law, Vineeth N. Balasubramanian et al.

The visual world naturally exhibits an imbalance in the number of object or scene instances resulting in a \emph{long-tailed distribution}. This imbalance poses significant challenges for classification models based on deep learning. Oversampling instances of the tail classes attempts to solve this imbalance. However, the limited visual diversity results in a network with poor representation ability. A simple counter to this is decoupling the representation and classifier networks and using oversampling only to train the classifier. In this paper, instead of repeatedly re-sampling the same image (and thereby features), we explore a direction that attempts to generate meaningful features by estimating the tail category's distribution. Inspired by ideas from recent work on few-shot learning, we create calibrated distributions to sample additional features that are subsequently used to train the classifier. Through several experiments on the CIFAR-100-LT (long-tail) dataset with varying imbalance factors and on mini-ImageNet-LT (long-tail), we show the efficacy of our approach and establish a new state-of-the-art. We also present a qualitative analysis of generated features using t-SNE visualizations and analyze the nearest neighbors used to calibrate the tail class distributions. Our code is available at https://github.com/rahulvigneswaran/TailCalibX.

Darshan Singh S, Anchit Gupta, C. V. Jawahar et al.

Over the last decade, online lecture videos have become increasingly popular and have experienced a meteoric rise during the pandemic. However, video-language research has primarily focused on instructional videos or movies, and tools to help students navigate the growing online lectures are lacking. Our first contribution is to facilitate research in the educational domain, by introducing AVLectures, a large-scale dataset consisting of 86 courses with over 2,350 lectures covering various STEM subjects. Each course contains video lectures, transcripts, OCR outputs for lecture frames, and optionally lecture notes, slides, assignments, and related educational content that can inspire a variety of tasks. Our second contribution is introducing video lecture segmentation that splits lectures into bite-sized topics that show promise in improving learner engagement. We formulate lecture segmentation as an unsupervised task that leverages visual, textual, and OCR cues from the lecture, while clip representations are fine-tuned on a pretext self-supervised task of matching the narration with the temporally aligned visual content. We use these representations to generate segments using a temporally consistent 1-nearest neighbor algorithm, TW-FINCH. We evaluate our method on 15 courses and compare it against various visual and textual baselines, outperforming all of them. Our comprehensive ablation studies also identify the key factors driving the success of our approach.

Balaji Darur, Amanmeet Garg, Makarand Tapaswi

Video Situation Recognition (VidSitu) addresses the challenging problem of "who did what to whom, with what, how, and where" in a video. It tests thorough video understanding by requiring identification of salient actions and associated short descriptions for event roles across multiple events. Grounding with VidSitu requires spatio-temporal localization of key entities across shots and varied appearances. We posit that coherent video understanding requires consistent identification of entities that play different roles. We propose Multimodal Entity Coreference (MEC) to unite entity descriptions in text with grounding across the video. Towards this, we introduce CineMEC, a multi-stage approach that unites event role mention groups with visual clusters of entities, without explicit grounding supervision during training. Our approach is designed to exploit the synergy between visual grounding and captioning, where improving one influences the other and vice versa. For evaluation, we extend the VidSitu dataset with grounding annotations. While previous work focuses primarily on descriptions, CineMEC improves consistency across both: captioning (+2.5% CIDEr, +7% LEA) and visual grounding (+18% HOTA).

Haran Raajesh, Naveen Reddy Desanur, Zeeshan Khan et al.

Characters are an important aspect of any storyline and identifying and including them in descriptions is necessary for story understanding. While previous work has largely ignored identity and generated captions with someone (anonymized names), recent work formulates id-aware captioning as a fill-in-the-blanks (FITB) task, where, given a caption with blanks, the goal is to predict person id labels. However, to predict captions with ids, a two-stage approach is required: first predict captions with someone, then fill in identities. In this work, we present a new single stage approach that can seamlessly switch between id-aware caption generation or FITB when given a caption with blanks. Our model, Movie-Identity Captioner (MICap), uses a shared auto-regressive decoder that benefits from training with FITB and full-caption generation objectives, while the encoder can benefit from or disregard captions with blanks as input. Another challenge with id-aware captioning is the lack of a metric to capture subtle differences between person ids. To this end, we introduce iSPICE, a caption evaluation metric that focuses on identity tuples created through intermediate scene graphs. We evaluate MICap on Large-Scale Movie Description Challenge (LSMDC), where we show a 4.2% improvement in FITB accuracy, and a 1-2% bump in classic captioning metrics.

Darshan Singh S, Zeeshan Khan, Makarand Tapaswi

While contrastive language image pretraining (CLIP) have exhibited impressive performance by learning highly semantic and generalized representations, recent works have exposed a fundamental drawback in its syntactic properties, that includes interpreting fine-grained attributes, actions, spatial relations, states, and details that require compositional reasoning. One reason for this is that natural captions often do not capture all the visual details of a scene. This leads to unaddressed visual concepts being misattributed to the wrong words. And the pooled image and text features, ends up acting as a bag of words, hence losing the syntactic information. In this work, we ask: Is it possible to enhance CLIP's fine-grained and syntactic abilities without compromising its semantic properties? We show that this is possible by adapting CLIP efficiently on a high-quality, comprehensive, and relatively small dataset. We demonstrate our adaptation strategy on VidSitu, a video situation recognition dataset annotated with verbs and rich semantic role labels (SRL). We use the SRL and verb information to create rule-based detailed captions, making sure they capture most of the visual concepts. Combined with hard negatives and hierarchical losses, these annotations allow us to learn a powerful visual representation, dubbed Fine-Grained CLIP (FiGCLIP), that preserves semantic understanding while being detail-oriented. We evaluate on five diverse vision-language tasks in both fine-tuning and zero-shot settings, achieving consistent improvements over the base CLIP model.

Piyush Bagad, Makarand Tapaswi, Cees G. M. Snoek et al.

We study the connection between audio-visual observations and the underlying physics of a mundane yet intriguing everyday activity: pouring liquids. Given only the sound of liquid pouring into a container, our objective is to automatically infer physical properties such as the liquid level, the shape and size of the container, the pouring rate and the time to fill. To this end, we: (i) show in theory that these properties can be determined from the fundamental frequency (pitch); (ii) train a pitch detection model with supervision from simulated data and visual data with a physics-inspired objective; (iii) introduce a new large dataset of real pouring videos for a systematic study; (iv) show that the trained model can indeed infer these physical properties for real data; and finally, (v) we demonstrate strong generalization to various container shapes, other datasets, and in-the-wild YouTube videos. Our work presents a keen understanding of a narrow yet rich problem at the intersection of acoustics, physics, and learning. It opens up applications to enhance multisensory perception in robotic pouring.

Aditya Kumar Singh, Dhruv Srivastava, Makarand Tapaswi

We introduce multimodal story summarization by leveraging TV episode recaps - short video sequences interweaving key story moments from previous episodes to bring viewers up to speed. We propose PlotSnap, a dataset featuring two crime thriller TV shows with rich recaps and long episodes of 40 minutes. Story summarization labels are unlocked by matching recap shots to corresponding sub-stories in the episode. We propose a hierarchical model TaleSumm that processes entire episodes by creating compact shot and dialog representations, and predicts importance scores for each video shot and dialog utterance by enabling interactions between local story groups. Unlike traditional summarization, our method extracts multiple plot points from long videos. We present a thorough evaluation on story summarization, including promising cross-series generalization. TaleSumm also shows good results on classic video summarization benchmarks.

Jona Ruthardt, Manu Gaur, Deva Ramanan et al.

Pretrained Vision Transformers (ViTs) such as DINOv2 and MAE provide generic image features that can be applied to a variety of downstream tasks such as retrieval, classification, and segmentation. However, such representations tend to focus on the most salient visual cues in the image, with no way to direct them toward less prominent concepts of interest. In contrast, Multimodal LLMs can be guided with textual prompts, but the resulting representations tend to be language-centric and lose their effectiveness for generic visual tasks. To address this, we introduce Steerable Visual Representations, a new class of visual representations, whose global and local features can be steered with natural language. While most vision-language models (e.g., CLIP) fuse text with visual features after encoding (late fusion), we inject text directly into the layers of the visual encoder (early fusion) via lightweight cross-attention. We introduce benchmarks for measuring representational steerability, and demonstrate that our steerable visual features can focus on any desired objects in an image while preserving the underlying representation quality. Our method also matches or outperforms dedicated approaches on anomaly detection and personalized object discrimination, exhibiting zero-shot generalization to out-of-distribution tasks.

Divy Kala, Eshika Khandelwal, Makarand Tapaswi

Audio descriptions (ADs) narrate important visual details in movies, enabling Blind and Low Vision (BLV) users to understand narratives and appreciate visual details. Existing works in automatic AD generation mostly focus on few-second trimmed clips, and evaluate them by comparing against a single ground-truth reference AD. However, writing ADs is inherently subjective. Through alignment and analysis of two independent AD tracks for the same movies, we quantify the subjectivity in when and whether to describe, and what and how to highlight. Thus, we show that working with trimmed clips is inadequate. We propose ADQA, a QA benchmark that evaluates ADs at the level of few-minute long, coherent video segments, testing whether they would help BLV users understand the story and appreciate visual details. ADQA features visual appreciation (VA) questions about visual facts and narrative understanding (NU) questions based on the plot. Through ADQA, we show that current AD generation methods lag far behind human-authored ADs. We conclude with several recommendations for future work and introduce a public leaderboard for benchmarking.

Darshana Saravanan, Makarand Tapaswi, Vineet Gandhi

To understand a prompt, Vision-Language models (VLMs) must perceive the image, comprehend the text, and build associations within and across both modalities. For instance, given an 'image of a red toy car', the model should associate this image to phrases like 'car', 'red toy', 'red object', etc. Feng and Steinhardt propose the Binding ID mechanism in LLMs, suggesting that the entity and its corresponding attribute tokens share a Binding ID in the model activations. We investigate this for image-text binding in VLMs using a synthetic dataset and task that requires models to associate 3D objects in an image with their descriptions in the text. Our experiments demonstrate that VLMs assign a distinct Binding ID to an object's image tokens and its textual references, enabling in-context association.

Eshika Khandelwal, Junyu Xie, Tengda Han et al.

Audio Descriptions (ADs) convey essential on-screen information, allowing visually impaired audiences to follow videos. To be effective, ADs must form a coherent sequence that helps listeners to visualise the unfolding scene, rather than describing isolated moments. However, most automatic methods generate each AD independently, often resulting in repetitive, incoherent descriptions. To address this, we propose a training-free method, CoherentAD, that first generates multiple candidate descriptions for each AD time interval, and then performs auto-regressive selection across the sequence to form a coherent and informative narrative. To evaluate AD sequences holistically, we introduce a sequence-level metric, StoryRecall, which measures how well the predicted ADs convey the ground truth narrative, alongside repetition metrics that capture the redundancy across consecutive AD outputs. Our method produces coherent AD sequences with enhanced narrative understanding, outperforming prior approaches that rely on independent generations.

Kawshik Manikantan, Shubham Toshniwal, Makarand Tapaswi et al.

The limited generalization of coreference resolution (CR) models has been a major bottleneck in the task's broad application. Prior work has identified annotation differences, especially for mention detection, as one of the main reasons for the generalization gap and proposed using additional annotated target domain data. Rather than relying on this additional annotation, we propose an alternative referential task, Major Entity Identification (MEI), where we: (a) assume the target entities to be specified in the input, and (b) limit the task to only the frequent entities. Through extensive experiments, we demonstrate that MEI models generalize well across domains on multiple datasets with supervised models and LLM-based few-shot prompting. Additionally, MEI fits the classification framework, which enables the use of robust and intuitive classification-based metrics. Finally, MEI is also of practical use as it allows a user to search for all mentions of a particular entity or a group of entities of interest.

Darshana Saravanan, Varun Gupta, Darshan Singh et al.

A fundamental aspect of compositional reasoning in a video is associating people and their actions across time. Recent years have seen great progress in general-purpose vision or video models and a move towards long-video understanding. While exciting, we take a step back and ask: are current models good at compositional reasoning on short videos? To this end, we introduce VELOCITI, a benchmark to study Video-LLMs by disentangling and assessing the comprehension of agents, actions, and their associations across multiple events. We adopt the Video-Language Entailment setup and propose StrictVLE that requires correct classification (rather than ranking) of the positive and negative caption. We evaluate several models and observe that even the best, LLaVA-OneVision (44.5%) and Gemini-1.5-Pro (49.3%), are far from human accuracy at 93.0%. Results show that action understanding lags behind agents, and negative captions created using entities appearing in the video perform worse than those obtained from pure text manipulation. We also present challenges with ClassicVLE and multiple-choice (MC) evaluation, strengthening our preference for StrictVLE. Finally, we validate that our benchmark requires visual inputs of multiple frames making it ideal to study video-language compositional reasoning.

Yash Khandelwal, Mayur Arvind, Sriram Kumar et al.

Malnutrition among newborns is a top public health concern in developing countries. Identification and subsequent growth monitoring are key to successful interventions. However, this is challenging in rural communities where health systems tend to be inaccessible and under-equipped, with poor adherence to protocol. Our goal is to equip health workers and public health systems with a solution for contactless newborn anthropometry in the community. We propose NurtureNet, a multi-task model that fuses visual information (a video taken with a low-cost smartphone) with tabular inputs to regress multiple anthropometry estimates including weight, length, head circumference, and chest circumference. We show that visual proxy tasks of segmentation and keypoint prediction further improve performance. We establish the efficacy of the model through several experiments and achieve a relative error of 3.9% and mean absolute error of 114.3 g for weight estimation. Model compression to 15 MB also allows offline deployment to low-cost smartphones.

Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi et al.

Following language instructions to navigate in unseen environments is a challenging problem for autonomous embodied agents. The agent not only needs to ground languages in visual scenes, but also should explore the environment to reach its target. In this work, we propose a dual-scale graph transformer (DUET) for joint long-term action planning and fine-grained cross-modal understanding. We build a topological map on-the-fly to enable efficient exploration in global action space. To balance the complexity of large action space reasoning and fine-grained language grounding, we dynamically combine a fine-scale encoding over local observations and a coarse-scale encoding on a global map via graph transformers. The proposed approach, DUET, significantly outperforms state-of-the-art methods on goal-oriented vision-and-language navigation (VLN) benchmarks REVERIE and SOON. It also improves the success rate on the fine-grained VLN benchmark R2R.

Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen et al.

Vision-and-language navigation (VLN) aims to enable embodied agents to navigate in realistic environments using natural language instructions. Given the scarcity of domain-specific training data and the high diversity of image and language inputs, the generalization of VLN agents to unseen environments remains challenging. Recent methods explore pretraining to improve generalization, however, the use of generic image-caption datasets or existing small-scale VLN environments is suboptimal and results in limited improvements. In this work, we introduce BnB, a large-scale and diverse in-domain VLN dataset. We first collect image-caption (IC) pairs from hundreds of thousands of listings from online rental marketplaces. Using IC pairs we next propose automatic strategies to generate millions of VLN path-instruction (PI) pairs. We further propose a shuffling loss that improves the learning of temporal order inside PI pairs. We use BnB pretrain our Airbert model that can be adapted to discriminative and generative settings and show that it outperforms state of the art for Room-to-Room (R2R) navigation and Remote Referring Expression (REVERIE) benchmarks. Moreover, our in-domain pretraining significantly increases performance on a challenging few-shot VLN evaluation, where we train the model only on VLN instructions from a few houses.

Vladimír Petrík, Makarand Tapaswi, Ivan Laptev et al.

Humans are adept at learning new tasks by watching a few instructional videos. On the other hand, robots that learn new actions either require a lot of effort through trial and error, or use expert demonstrations that are challenging to obtain. In this paper, we explore a method that facilitates learning object manipulation skills directly from videos. Leveraging recent advances in 2D visual recognition and differentiable rendering, we develop an optimization based method to estimate a coarse 3D state representation for the hand and the manipulated object(s) without requiring any supervision. We use these trajectories as dense rewards for an agent that learns to mimic them through reinforcement learning. We evaluate our method on simple single- and two-object actions from the Something-Something dataset. Our approach allows an agent to learn actions from single videos, while watching multiple demonstrations makes the policy more robust. We show that policies learned in a simulated environment can be easily transferred to a real robot.

Vivek Sharma, Makarand Tapaswi, Rainer Stiefelhagen

True understanding of videos comes from a joint analysis of all its modalities: the video frames, the audio track, and any accompanying text such as closed captions. We present a way to learn a compact multimodal feature representation that encodes all these modalities. Our model parameters are learned through a proxy task of inferring the temporal ordering of a set of unordered videos in a timeline. To this end, we create a new multimodal dataset for temporal ordering that consists of approximately 30K scenes (2-6 clips per scene) based on the "Large Scale Movie Description Challenge". We analyze and evaluate the individual and joint modalities on three challenging tasks: (i) inferring the temporal ordering of a set of videos; and (ii) action recognition. We demonstrate empirically that multimodal representations are indeed complementary, and can play a key role in improving the performance of many applications.

Vivek Sharma, Makarand Tapaswi, M. Saquib Sarfraz et al.

A good clustering algorithm can discover natural groupings in data. These groupings, if used wisely, provide a form of weak supervision for learning representations. In this work, we present Clustering-based Contrastive Learning (CCL), a new clustering-based representation learning approach that uses labels obtained from clustering along with video constraints to learn discriminative face features. We demonstrate our method on the challenging task of learning representations for video face clustering. Through several ablation studies, we analyze the impact of creating pair-wise positive and negative labels from different sources. Experiments on three challenging video face clustering datasets: BBT-0101, BF-0502, and ACCIO show that CCL achieves a new state-of-the-art on all datasets.

Anna Kukleva, Makarand Tapaswi, Ivan Laptev

Interactions between people are often governed by their relationships. On the flip side, social relationships are built upon several interactions. Two strangers are more likely to greet and introduce themselves while becoming friends over time. We are fascinated by this interplay between interactions and relationships, and believe that it is an important aspect of understanding social situations. In this work, we propose neural models to learn and jointly predict interactions, relationships, and the pair of characters that are involved. We note that interactions are informed by a mixture of visual and dialog cues, and present a multimodal architecture to extract meaningful information from them. Localizing the pair of interacting characters in video is a time-consuming process, instead, we train our model to learn from clip-level weak labels. We evaluate our models on the MovieGraphs dataset and show the impact of modalities, use of longer temporal context for predicting relationships, and achieve encouraging performance using weak labels as compared with ground-truth labels. Code is online.

Atef Chaudhury, Makarand Tapaswi, Seung Wook Kim et al.

Understanding stories is a challenging reading comprehension problem for machines as it requires reading a large volume of text and following long-range dependencies. In this paper, we introduce the Shmoop Corpus: a dataset of 231 stories that are paired with detailed multi-paragraph summaries for each individual chapter (7,234 chapters), where the summary is chronologically aligned with respect to the story chapter. From the corpus, we construct a set of common NLP tasks, including Cloze-form question answering and a simplified form of abstractive summarization, as benchmarks for reading comprehension on stories. We then show that the chronological alignment provides a strong supervisory signal that learning-based methods can exploit leading to significant improvements on these tasks. We believe that the unique structure of this corpus provides an important foothold towards making machine story comprehension more approachable.

Makarand Tapaswi, Marc T. Law, Sanja Fidler

Understanding videos such as TV series and movies requires analyzing who the characters are and what they are doing. We address the challenging problem of clustering face tracks based on their identity. Different from previous work in this area, we choose to operate in a realistic and difficult setting where: (i) the number of characters is not known a priori; and (ii) face tracks belonging to minor or background characters are not discarded. To this end, we propose Ball Cluster Learning (BCL), a supervised approach to carve the embedding space into balls of equal size, one for each cluster. The learned ball radius is easily translated to a stopping criterion for iterative merging algorithms. This gives BCL the ability to estimate the number of clusters as well as their assignment, achieving promising results on commonly used datasets. We also present a thorough discussion of how existing metric learning literature can be adapted for this task.

Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac et al.

Learning text-video embeddings usually requires a dataset of video clips with manually provided captions. However, such datasets are expensive and time consuming to create and therefore difficult to obtain on a large scale. In this work, we propose instead to learn such embeddings from video data with readily available natural language annotations in the form of automatically transcribed narrations. The contributions of this work are three-fold. First, we introduce HowTo100M: a large-scale dataset of 136 million video clips sourced from 1.22M narrated instructional web videos depicting humans performing and describing over 23k different visual tasks. Our data collection procedure is fast, scalable and does not require any additional manual annotation. Second, we demonstrate that a text-video embedding trained on this data leads to state-of-the-art results for text-to-video retrieval and action localization on instructional video datasets such as YouCook2 or CrossTask. Finally, we show that this embedding transfers well to other domains: fine-tuning on generic Youtube videos (MSR-VTT dataset) and movies (LSMDC dataset) outperforms models trained on these datasets alone. Our dataset, code and models will be publicly available at: www.di.ens.fr/willow/research/howto100m/.

Vivek Sharma, Makarand Tapaswi, M. Saquib Sarfraz et al.

Analyzing the story behind TV series and movies often requires understanding who the characters are and what they are doing. With improving deep face models, this may seem like a solved problem. However, as face detectors get better, clustering/identification needs to be revisited to address increasing diversity in facial appearance. In this paper, we address video face clustering using unsupervised methods. Our emphasis is on distilling the essential information, identity, from the representations obtained using deep pre-trained face networks. We propose a self-supervised Siamese network that can be trained without the need for video/track based supervision, and thus can also be applied to image collections. We evaluate our proposed method on three video face clustering datasets. The experiments show that our methods outperform current state-of-the-art methods on all datasets. Video face clustering is lacking a common benchmark as current works are often evaluated with different metrics and/or different sets of face tracks.

Seung Wook Kim, Makarand Tapaswi, Sanja Fidler

Humans learn to solve tasks of increasing complexity by building on top of previously acquired knowledge. Typically, there exists a natural progression in the tasks that we learn - most do not require completely independent solutions, but can be broken down into simpler subtasks. We propose to represent a solver for each task as a neural module that calls existing modules (solvers for simpler tasks) in a functional program-like manner. Lower modules are a black box to the calling module, and communicate only via a query and an output. Thus, a module for a new task learns to query existing modules and composes their outputs in order to produce its own output. Our model effectively combines previous skill-sets, does not suffer from forgetting, and is fully differentiable. We test our model in learning a set of visual reasoning tasks, and demonstrate improved performances in all tasks by learning progressively. By evaluating the reasoning process using human judges, we show that our model is more interpretable than an attention-based baseline.

Paul Vicol, Makarand Tapaswi, Lluis Castrejon et al.

There is growing interest in artificial intelligence to build socially intelligent robots. This requires machines to have the ability to "read" people's emotions, motivations, and other factors that affect behavior. Towards this goal, we introduce a novel dataset called MovieGraphs which provides detailed, graph-based annotations of social situations depicted in movie clips. Each graph consists of several types of nodes, to capture who is present in the clip, their emotional and physical attributes, their relationships (i.e., parent/child), and the interactions between them. Most interactions are associated with topics that provide additional details, and reasons that give motivations for actions. In addition, most interactions and many attributes are grounded in the video with time stamps. We provide a thorough analysis of our dataset, showing interesting common-sense correlations between different social aspects of scenes, as well as across scenes over time. We propose a method for querying videos and text with graphs, and show that: 1) our graphs contain rich and sufficient information to summarize and localize each scene; and 2) subgraphs allow us to describe situations at an abstract level and retrieve multiple semantically relevant situations. We also propose methods for interaction understanding via ordering, and reason understanding. MovieGraphs is the first benchmark to focus on inferred properties of human-centric situations, and opens up an exciting avenue towards socially-intelligent AI agents.

Ruiyu Li, Makarand Tapaswi, Renjie Liao et al.

We address the problem of recognizing situations in images. Given an image, the task is to predict the most salient verb (action), and fill its semantic roles such as who is performing the action, what is the source and target of the action, etc. Different verbs have different roles (e.g. attacking has weapon), and each role can take on many possible values (nouns). We propose a model based on Graph Neural Networks that allows us to efficiently capture joint dependencies between roles using neural networks defined on a graph. Experiments with different graph connectivities show that our approach that propagates information between roles significantly outperforms existing work, as well as multiple baselines. We obtain roughly 3-5% improvement over previous work in predicting the full situation. We also provide a thorough qualitative analysis of our model and influence of different roles in the verbs.

Manuel Martinez, Monica Haurilet, Ziad Al-Halah et al.

The Earth Mover's Distance (EMD) computes the optimal cost of transforming one distribution into another, given a known transport metric between them. In deep learning, the EMD loss allows us to embed information during training about the output space structure like hierarchical or semantic relations. This helps in achieving better output smoothness and generalization. However EMD is computationally expensive.Moreover, solving EMD optimization problems usually require complex techniques like lasso. These properties limit the applicability of EMD-based approaches in large scale machine learning. We address in this work the difficulties facing incorporation of EMD-based loss in deep learning frameworks. Additionally, we provide insight and novel solutions on how to integrate such loss function in training deep neural networks. Specifically, we make three main contributions: (i) we provide an in-depth analysis of the fastest state-of-the-art EMD algorithm (Sinkhorn Distance) and discuss its limitations in deep learning scenarios. (ii) we derive fast and numerically stable closed-form solutions for the EMD gradient in output spaces with chain- and tree- connectivity; and (iii) we propose a relaxed form of the EMD gradient with equivalent computational complexity but faster convergence rate. We support our claims with experiments on real datasets. In a restricted data setting on the ImageNet dataset, we train a model to classify 1000 categories using 50K images, and demonstrate that our relaxed EMD loss achieves better Top-1 accuracy than the cross entropy loss. Overall, we show that our relaxed EMD loss criterion is a powerful asset for deep learning in the small data regime.

Ziad Al-Halah, Makarand Tapaswi, Rainer Stiefelhagen

Collecting training images for all visual categories is not only expensive but also impractical. Zero-shot learning (ZSL), especially using attributes, offers a pragmatic solution to this problem. However, at test time most attribute-based methods require a full description of attribute associations for each unseen class. Providing these associations is time consuming and often requires domain specific knowledge. In this work, we aim to carry out attribute-based zero-shot classification in an unsupervised manner. We propose an approach to learn relations that couples class embeddings with their corresponding attributes. Given only the name of an unseen class, the learned relationship model is used to automatically predict the class-attribute associations. Furthermore, our model facilitates transferring attributes across data sets without additional effort. Integrating knowledge from multiple sources results in a significant additional improvement in performance. We evaluate on two public data sets: Animals with Attributes and aPascal/aYahoo. Our approach outperforms state-of-the-art methods in both predicting class-attribute associations and unsupervised ZSL by a large margin.

Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen et al.

We introduce the MovieQA dataset which aims to evaluate automatic story comprehension from both video and text. The dataset consists of 14,944 questions about 408 movies with high semantic diversity. The questions range from simpler "Who" did "What" to "Whom", to "Why" and "How" certain events occurred. Each question comes with a set of five possible answers; a correct one and four deceiving answers provided by human annotators. Our dataset is unique in that it contains multiple sources of information -- video clips, plots, subtitles, scripts, and DVS. We analyze our data through various statistics and methods. We further extend existing QA techniques to show that question-answering with such open-ended semantics is hard. We make this data set public along with an evaluation benchmark to encourage inspiring work in this challenging domain.