Visual Planning for Assistance (VPA) aims to predict a sequence of user actions required to achieve a specified goal based on a video showing the user’s progress. Although recent advances in multimodal large language models (MLLMs) have shown promising results in video understanding, long-horizon visual planning remains a challenging problem. We identify two challenges in training large MLLMs for video-based planning tasks: (1) scarcity of procedural annotations, limiting the model’s ability to learn procedural task dynamics effectively, and (2) inefficiency of next-token prediction objective to explicitly capture the structured action space for visual planning when compared to free-form, natural language. To tackle data scarcity, we introduce Auxiliary Task Augmentation. We design and train our model on auxiliary tasks relevant to long-horizon video-based planning (e.g., goal prediction) to augment the model’s planning ability. To more explicitly model the structured action space unique to visual planning tasks, we leverage Multi-token Prediction, extending traditional next-token prediction by using multiple heads to predict multiple future tokens during training. Our approach, VideoPlan, achieves state-of-the-art VPA performance on the COIN and CrossTask datasets, surpassing prior methods by 7.3% and 3.4%, respectively, when predicting 3 future actions. We further extend our method to the challenging Ego4D Long-term Action Anticipation task, and show that it is on par with the state-of-the-art approaches despite not using specialized egocentric features. Code is available at https://github.com/CeeZh/VideoPlan.

There has recently been a surge of interest in Wearable Assistant Agents: agents embodied in a wearable form factor such as smart glasses, who can take actions toward a user’s stated goal — a high-level language-expressed command such as “where did I leave my keys?”, “Text Alice I will be late”, or “What’s the weather in Cancun?”. In this work, we consider the complementary problem of eliminating the effort required to interact with such an agent by proactively inferring the user’s goal from multimodal contextual observations. As vision-language models (VLMs) hold strong potential to ultimately solve this problem, our work focuses on creating a strong benchmark to measure progress toward this end. Given the limited prior work in this area, establishing the benchmark required collecting a novel multimodal goal-inference dataset; our dataset comprises ~30 hours of data from 363 participants across 3,482 recordings, featuring ground-truth reference goals alongside accompanying visual, audio, digital, and longitudinal contextual observations. We ran a human predictability study, where we found that humans set a strong baseline that comprises a de facto upper bound on model performance: they show multiple choice question (MCQ) accuracy of 93%, with the best VLM achieving about 84% accuracy. However, MCQ assesses discrimination, not the model’s ultimate task of generating the goal through open-ended text generation. Through a meta-evaluation, we find that a VLM judging the generated goals is as good as a human judge if it has access to a human-authored script of the video or a correct reference goal. Finally, we evaluate several families of modern vision-language models on the benchmark, showing that larger models have a significant performance advantage, but are still far from being practically useful, as they produce relevant goals only ~57% of the time. The best-performing smaller models—whose size makes them better suited to wearable applications—perform significantly worse than their counterparts, generating ~49% accuracy on the benchmark. Through a modality ablation, we show that models benefit from extra information in relevant modalities with minimal performance degradation from irrelevant modalities, but don’t gain as much when noisy modalities are included (e.g., in the case of digital context when most of the app state is irrelevant).

In deep metric learning for visual recognition, the calibration of distance thresholds is crucial for achieving desired model performance in the true positive rates (TPR) or true negative rates (TNR). However, calibrating this threshold presents challenges in open-world scenarios, where the test classes can be entirely disjoint from those encountered during training. We define the problem of finding distance thresholds for a trained embedding model to achieve target performance metrics over unseen open-world test classes as open-world threshold calibration. Existing posthoc threshold calibration methods, reliant on inductive inference and requiring a calibration dataset with a similar distance distribution as the test data, often prove ineffective in open-world scenarios. To address this, we introduce OpenGCN, a Graph Neural Network-based transductive threshold calibration method with enhanced adaptability and robustness. OpenGCN learns to predict pairwise connectivity for the unlabeled test instances embedded in a graph to determine its TPR and TNR at various distance thresholds, allowing for transductive inference of the distance thresholds which also incorporates test-time information. Extensive experiments across open-world visual recognition benchmarks validate OpenGCN's superiority over existing posthoc calibration methods for open-world threshold calibration.

Existing losses used in deep metric learning (DML) for image retrieval often lead to highly non-uniform intra-class and inter-class representation structures across test classes and data distributions. When combined with the common practice of using a fixed threshold to declare a match, this gives rise to significant performance variations in terms of false accept rate (FAR) and false reject rate (FRR) across test classes and data distributions. We define this issue in DML as threshold inconsistency. In real-world applications, such inconsistency often complicates the threshold selection process when deploying large-scale image retrieval systems. To measure this inconsistency, we propose a novel variance-based metric called Operating-Point-Inconsistency-Score (OPIS) that quantifies the variance in the operating characteristics across classes. Using the OPIS metric, we find that achieving high accuracy levels in a DML model does not automatically guarantee threshold consistency. In fact, our investigation reveals a Pareto frontier in the high-accuracy regime, where existing methods to improve accuracy often lead to degradation in threshold consistency. To address this trade-off, we introduce the Threshold-Consistent Margin (TCM) loss, a simple yet effective regularization technique that promotes uniformity in representation structures across classes by selectively penalizing hard sample pairs. Large-scale experiments demonstrate TCM's effectiveness in enhancing threshold consistency while preserving accuracy, simplifying the threshold selection process in practical DML settings.

Multi-dataset training provides a viable solution for exploiting heterogeneous large-scale datasets without extra annotation cost. In this work, we propose a scalable multi-dataset detector (ScaleDet) that can scale up its generalization across datasets when increasing the number of training datasets. Unlike existing multi-dataset learners that mostly rely on manual relabelling efforts or sophisticated optimizations to unify labels across datasets, we introduce a simple yet scalable formulation to derive a unified semantic label space for multi-dataset training. ScaleDet is trained by visual-textual alignment to learn the label assignment with label semantic similarities across datasets. Once trained, ScaleDet can generalize well on any given upstream and downstream datasets with seen and unseen classes. We conduct extensive experiments using LVIS, COCO, Objects365, OpenImages as upstream datasets, and 13 datasets from Object Detection in the Wild (ODinW) as downstream datasets. Our results show that ScaleDet achieves compelling strong model performance with an mAP of 50.7 on LVIS, 58.8 on COCO, 46.8 on Objects365, 76.2 on OpenImages, and 71.8 on ODinW, surpassing state-of-the-art detectors with the same backbone.

We present SkeleTR, a new framework for skeleton-based action recognition. In contrast to prior work, which focuses mainly on controlled environments, we target more general scenarios that typically involve a variable number of people and various forms of interaction between people. SkeleTR works with a two-stage paradigm. It first models the intra-person skeleton dynamics for each skeleton sequence with graph convolutions, and then uses stacked Transformer encoders to capture person interactions that are important for action recognition in general scenarios. To mitigate the negative impact of inaccurate skeleton associations, SkeleTR takes relative short skeleton sequences as input and increases the number of sequences. As a unified solution, SkeleTR can be directly applied to multiple skeleton-based action tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It also enables transfer learning and joint training across different action tasks and datasets, which result in performance improvement. When evaluated on various skeleton-based action recognition benchmarks, SkeleTR achieves the state-of-the-art performance.

We propose a novel one-stage Transformer-based semantic and spatial refined transformer (SSRT) to solve the Human-Object Interaction detection task, which requires to localize humans and objects, and predicts their interactions. Differently from previous Transformer-based HOI approaches, which mostly focus at improving the design of the decoder outputs for the final detection, SSRT introduces two new modules to help select the most relevant object-action pairs within an image and refine the queries' representation using rich semantic and spatial features. These enhancements lead to state-of-the-art results on the two most popular HOI benchmarks: V-COCO and HICO-DET.

Learning a unified person detection and re-identification model is a key component of modern trackers. However, training such models usually relies on the availability of training images / videos that are manually labeled with both person boxes and their identities. In this work, we explore training such a model by only using person box annotations, thus removing the necessity of manually labeling a training dataset with additional person identity annotation as these are expensive to collect. To this end, we present a contrastive learning framework to learn person similarity without using manually labeled identity annotations. First, we apply image-level augmentation to images on public person detection datasets, based on which we learn a strong model for general person detection as well as for short-term person re-identification. To learn a model capable of longerterm re-identification, we leverage the natural appearance evolution of each person in videos to serve as instance-level appearance augmentation in our contrastive loss formulation. Without access to the target dataset or person identity annotation, our model achieves competitive results compared to existing fully-supervised state-of-the-art methods on both person search and person tracking tasks. Our model also shows promising results for saving the annotation cost that is needed to achieve a certain level of performance on the person search task.

Most self-supervised video representation learning approaches focus on action recognition. In contrast, in this paper we focus on self-supervised video learning for movie understanding and propose a novel hierarchical self-supervised pretraining strategy that separately pretrains each level of our hierarchical movie understanding model (based on [37]). Specifically, we propose to pretrain the low-level video backbone using a contrastive learning objective, while pretrain the higher-level video contextualizer using an event mask prediction task, which enables the usage of different data sources for pretraining different levels of the hierarchy. We first show that our self-supervised pre-training strategies are effective and lead to improved performance on all tasks and metrics on VidSitu benchmark [37] (e.g., improving on semantic role prediction from 47% to 61% CIDEr scores). We further demonstrate the effectiveness of our contextualized event features on LVU tasks [54], both when used alone and when combined with instance features, showing their complementarity.

We propose TubeR: a simple solution for spatio-temporal video action detection. Different from existing methods that depend on either an offline actor detector or hand-designed actor-positional hypotheses like proposals or anchors, we propose to directly detect an action tubelet in a video by simultaneously performing action localization and recognition from a single representation. TubeR learns a set of tubelet-queries and utilizes a tubelet-attention module to model the dynamic spatio-temporal nature of a video clip, which effectively reinforces the model capacity compared to using actor-positional hypotheses in the spatio-temporal space. For videos containing transitional states or scene changes, we propose a context aware classification head to utilize short-term and long-term context to strengthen action classification, and an action switch regression head for detecting the precise temporal action extent. TubeR directly produces action tubelets with variable lengths and even maintains good results for long video clips. TubeR outperforms the previous state-of-the-art on commonly used action detection datasets AVA, UCF101-24 and JHMDB51-21. Code will be available on GluonCV(https://cv.gluon.ai/).

In this paper, we provide an in-depth study of Stochastic Backpropagation (SBP) when training deep neural networks for standard image classification and object detection tasks. During backward propagation, SBP calculates gradients by using only a subset of feature maps to save GPU memory and computational cost. We interpret SBP as an efficient way to implement stochastic gradient decent by performing backpropagation dropout, which leads to significant memory saving and training run-time reduction, with a minimal impact on the overall model accuracy. We offer best practices to apply SBP for training image recognition models, which can be adopted in learning a wide range of deep neural networks. Experiments on image classification and object detection show that SBP can save up to 40% of GPU memory with less than 1% accuracy degradation. Code is available at: https://github.com/amazon-research/stochastic-backpropagation

In the world of action recognition research, one primary focus has been on how to construct and train networks to model the spatial-temporal volume of an input video. These methods typically uniformly sample a segment of an input clip (along the temporal dimension). However, not all parts of a video are equally important to determine the action in the clip. In this work, we focus instead on learning where to extract features, so as to focus on the most informative parts of the video. We propose a method called the non-uniform temporal aggregation (NUTA), which aggregates features only from informative temporal segments. We also introduce a synchronization method that allows our NUTA features to be temporally aligned with traditional uniformly sampled video features, so that both local and clip-level features can be combined. Our model has achieved state-of-the-art performance on four widely used large-scale action-recognition datasets (Kinetics400, Kinetics700, Something-something V2 and Charades). In addition, we have created a visualization to illustrate how the proposed NUTA method selects only the most relevant parts of a video clip.

Temporal action segmentation is a task to classify each frame in the video with an action label. However, it is quite expensive to annotate every frame in a large corpus of videos to construct a comprehensive supervised training dataset. Thus in this work we propose an unsupervised method, namely SSCAP, that operates on a corpus of unlabeled videos and predicts a likely set of temporal segments across the videos. SSCAP leverages Self-Supervised learning to extract distinguishable features and then applies a novel Co-occurrence Action Parsing algorithm to not only capture the correlation among sub-actions under-lying the structure of activities, but also estimate the temporal path of the sub-actions in an accurate and general way. We evaluate on both classic datasets (Breakfast, 50Sal-ads) and the emerging fine-grained action dataset (Fine-Gym) with more complex activity structures and similar sub-actions. Results show that SSCAP achieves state-of-the-art performance on all datasets and can even outperform some weakly-supervised approaches, demonstrating its effectiveness and generalizability.

In this paper, we focus on improving online multi-object tracking (MOT). In particular, we introduce a region-based Siamese Multi-Object Tracking network, which we name SiamMOT. SiamMOT includes a motion model that estimates the instance’s movement between two frames such that detected instances are associated. To explore how the motion modelling affects its tracking capability, we present two variants of Siamese tracker, one that implicitly models motion and one that models it explicitly. We carry out extensive quantitative experiments on three different MOT datasets: MOT17, TAO-person and Caltech Roadside Pedestrians, showing the importance of motion modelling for MOT and the ability of SiamMOT to substantially outperform the state-of-the-art. Finally, SiamMOT also outperforms the winners of ACM MM’20 HiEve Grand Challenge on HiEve dataset. Moreover, SiamMOT is efficient, and it runs at 17 FPS for 720P videos on a single modern GPU.

We propose a fully automated system that simultaneously estimates the camera intrinsics, the ground plane, and physical distances between people from a single RGB image or video captured by a camera viewing a 3-D scene from a fixed vantage point. To automate camera calibration and distance estimation, we leverage priors about human pose and develop a novel direct formulation for pose-based auto-calibration and distance estimation, which shows state-of-the-art performance on publicly available datasets. The proposed approach enables existing camera systems to measure physical distances without needing a dedicated calibration process or range sensors, and is applicable to a broad range of use cases such as social distancing and workplace safety. Furthermore, to enable evaluation and drive research in this area, we contribute to the publicly available MEVA dataset with additional distance annotations, resulting in "MEVADA" – an evaluation benchmark for the pose-based auto-calibration and distance estimation problem.

Most action recognition solutions rely on dense sampling to precisely cover the informative temporal clip. Extensively searching temporal region is expensive for a real-world application. In this work, we focus on improving the inference efficiency of current action recognition backbones on trimmed videos, and illustrate that an action model can accurately classify an action with a single pass over the video unlike the multi-clip sampling common with SOTA by learning to drop non-informative features. We present Selective Feature Compression (SFC), an action recognition inference strategy that greatly increases model inference efficiency without compromising accuracy. Different from previous works that compress kernel size and decrease the channel dimension, we propose to compress features along the spatio-temporal dimensions without the need to change backbone parameters. Our experiments on Kinetics-400, UCF101 and ActivityNet show that SFC is able to reduce inference speed by 6-7x and memory usage by 5-6x compared with the commonly used 30 crop dense sampling procedure, while also slightly improving Top1 Accuracy. We perform thorough quantitative and qualitative evaluation and show how our SFC learns to attend to important video regions for the task of action recognition.

We introduce Video Transformer (VidTr) with separable-attention for video classification. Comparing with commonly used 3D networks, VidTr is able to aggregate spatiotemporal information via stacked attentions and provide better performance with higher efficiency. We first introduce the vanilla video transformer and show that transformer module is able to perform spatio-temporal modeling from raw pixels, but with heavy memory usage. We then present VidTr which reduces the memory cost by 3.3× while keeping the same performance. To further optimize the model, we propose the standard deviation based topK pooling for attention (pooltopK_std), which reduces the computation by dropping non-informative features along temporal dimension. VidTr achieves state-of-the-art performance on five commonly used datasets with lower computational requirement, showing both the efficiency and effectiveness of our design. Finally, error analysis and visualization show that VidTr is especially good at predicting actions that require long-term temporal reasoning.

In crowd counting we often observe wrong predictions on image regions not containing any person. But how often do these mistakes happen and how much do they affect the overall performance? In this paper we analyze this problem in depth and present an extensive analysis on five of the most important crowd counting datasets. We present this analysis in two parts. First, we quantify the number of mistakes made. Our results show that (i) mistakes on back-ground are substantial and they are responsible for 18-49% of the total error, (ii) models do not generalize well to different kinds of backgrounds and perform poorly on completely background images, and (iii) models make many more mistakes than those captured by the standard Mean Absolute Error (MAE) metric, as counting on background compensates substantially for misses on foreground. And second, we quantify the performance change gained by helping the model better deal with this problem. We enrich a popular crowd counting network with a segmentation branch trained to suppress background predictions. This simple addition (i) reduces background error by 10-83%, (ii) reduces fore-ground error by up to 26% and (iii) improves overall crowd counting performance up to 20%. When compared against the literature, this simple technique achieves very competitive results on all datasets, showing the importance of tack-ling the background problem.

Trained on large datasets, deep learning (DL) can accurately classify videos into hundreds of diverse classes. However, video data is expensive to annotate. Zero-shot learning (ZSL) proposes one solution to this problem. ZSL trains a model once, and generalizes to new tasks whose classes are not present in the training dataset. We propose the first end-to-end algorithm for ZSL in video classification. Our training procedure builds on insights from recent video classification literature and uses a trainable 3D CNN to learn the visual features. This is in contrast to previous video ZSL methods, which use pretrained feature extractors. We also extend the current benchmarking paradigm: Previous techniques aim to make the test task unknown at training time but fall short of this goal. We encourage domain shift across training and test data and disallow tailoring a ZSL model to a specific test dataset. We outperform the state-of-the-art by a wide margin. Our code, evaluation procedure and model weights are available online github.com/bbrattoli/ZeroShotVideoClassification.

We propose a novel top-down approach that tackles the problem of multi-person human pose estimation and tracking in videos. In contrast to existing top-down approaches, our method is not limited by the performance of its person detector and can predict the poses of person instances not localized. It achieves this capability by propagating known person locations forward and backward in time and searching for poses in those regions. Our approach consists of three components: (i) a Clip Tracking Network that performs body joint detection and tracking simultaneously on small video clips; (ii) a Video Tracking Pipeline that merges the fixed-length tracklets produced by the Clip Tracking Network to arbitrary length tracks; and (iii) a Spatial-Temporal Merging procedure that refines the joint locations based on spatial and temporal smoothing terms. Thanks to the precision of our Clip Tracking Network and our merging procedure, our approach produces very accurate joint predictions and can fix common mistakes on hard scenarios like heavily entangled people. Our approach achieves state-of-the-art results on both joint detection and tracking, on both the PoseTrack 2017 and 2018 datasets, and against all top-down and bottom-down approaches.

Multi-object tracking systems often consist of a combination of a detector, a short term linker, a re-identification feature extractor and a solver that takes the output from these separate components and makes a final prediction. Differently, this work aims to unify all these in a single tracking system. Towards this, we propose Siamese Track-RCNN, a two stage detect-and-track framework which consists of three functional branches: (1) the detection branch localizes object instances; (2) the Siamese-based track branch estimates the object motion and (3) the object re-identification branch re-activates the previously terminated tracks when they re-emerge. We used this design and apply it to the Human in Events dataset.

As classifications datasets progressively get larger in terms of label space and number of examples, annotating them with all labels becomes non-trivial and expensive task. For example, annotating the entire OpenImage test set can cost $6.5M. Hence, in current large-scale benchmarks such as OpenImages and LVIS, less than 1\% of the labels are annotated across all images. Standard classification models are trained in a manner where these un-annotated labels are ignored. Ignoring these un-annotated labels result in loss of supervisory signal which reduces the performance of the classification models. Instead, in this paper, we exploit relationships among images and labels to derive more supervisory signal from the un-annotated labels. We study the effectiveness of our approach across several multi-label computer vision benchmarks, such as CIFAR100, MS-COCO panoptic segmentation, OpenImage and LVIS datasets. Our approach can outperform baselines by a margin of 2-10% across all the datasets on mean average precision (mAP) and mean F1 metrics.

Action recognition has seen a dramatic performance improvement in the last few years. Most of the current state-of-the-art literature either aims at improving performance through changes to the backbone CNN network, or exploring different trade-offs between computational efficiency and performance, again through altering the backbone network. However, almost all of these works maintain the same last layers of the network, which simply consist of a global average pooling followed by a fully connected layer. In this work we focus on how to improve the representation capacity of the network, but rather than altering the backbone, we focus on improving the last layers of the network, where changes have low impact in terms of computational cost. In particular, we hypothesize that current architectures have poor sensitivity to finer details and we exploit recent advances in the fine-grained recognition literature to improve our model in this aspect. With the proposed approach, we obtain state-of-the-art performance on Kinetics-400 and Something-Something-V1, the two major large-scale action recognition benchmarks.

This work proposes a method to interpret a scene by assigning a semantic label at every pixel and inferring the spatial extent of individual object instances together with their occlusion relationships. Starting with an initial pixel labeling and a set of candidate object masks for a given test image, we select a subset of objects that explain the image well and have valid overlap relationships and occlusion ordering. This is done by minimizing an integer quadratic program either using a greedy method or a standard solver. Then we alternate between using the object predictions to refine the pixel labels and vice versa. The proposed system obtains promising results on two challenging subsets of the LabelMe and SUN datasets, the largest of which contains 45, 676 images and 232 classes.

We consider the problem of estimating the relative depth of a scene from a monocular image. The dark channel prior, used as a statistical observation of haze free images, has been previously leveraged for haze removal and relative depth estimation tasks. However, as a local measure, it fails to account for higher order semantic relationship among scene elements. We propose a dual channel prior used for identifying pixels that are unlikely to comply with the dark channel assumption, leading to erroneous depth estimates. We further leverage semantic segmentation information and patch match label propagation to enforce semantically consistent geometric priors. Experiments illustrate the quantitative and qualitative advantages of our approach when compared to state of the art methods.

This paper presents a system for image parsing, or labeling each pixel in an image with its semantic category, aimed at achieving broad coverage across hundreds of object categories, many of them sparsely sampled. The system combines region-level features with per-exemplar sliding window detectors. Per-exemplar detectors are better suited for our parsing task than traditional bounding box detectors: they perform well on classes with little training data and high intra-class variation, and they allow object masks to be transferred into the test image for pixel-level segmentation. The proposed system achieves state-of-the-art accuracy on three challenging datasets, the largest of which contains 45,676 images and 232 labels.

In this work, we address the issue of geometric verification, with a focus on modeling large-scale landmark image collections gathered from the internet. In particular, we show that we can compute and learn descriptive statistics pertaining to the image collection by leveraging information that arises as a by-product of the matching and verification stages. Our approach is based on the intuition that validating numerous image pairs of the same geometric scene structures quickly reveals useful information about two aspects of the image collection: (a) the reliability of individual visual words and (b) the appearance of landmarks in the image collection. Both of these sources of information can then be used to drive any subsequent processing, thus allowing the system to bootstrap itself. While current techniques make use of dedicated training/preprocessing stages, our approach elegantly integrates into the standard geometric verification pipeline, by simply leveraging the information revealed during the verification stage. The main result of this work is that this unsupervised “learning-as-you-go ” approach significantly improves performance; our experiments demonstrate significant improvements in efficiency and completeness over standard techniques.

This paper presents a framework for image parsing with multiple label sets. For example, we may want to simultaneously label every image region according to its basic-level object category (car, building, road, tree, etc.), superordinate category (animal, vehicle, manmade object, natural object, etc.), geometric orientation (horizontal, vertical, etc.), and material (metal, glass, wood, etc.). Some object regions may also be given part names (a car can have wheels, doors, windshield, etc.). We compute co-occurrence statistics between different label types of the same region to capture relationships such as “roads are horizontal,” “cars are made of metal,” “cars have wheels” but “horses have legs,” and so on. By incorporating these constraints into a Markov Random Field inference framework and jointly solving for all the label sets, we are able to improve the classification accuracy for all the label sets at once, achieving a richer form of image understanding.