In this paper, we propose VidLA, an approach for video-language alignment at scale. There are two major limitations of previous video-language alignment approaches. First, they do not capture both short-range and long-range temporal dependencies and typically employ complex hierarchical deep network architectures that are hard to integrate with existing pretrained image-text foundation models. To effectively address this limitation, we instead keep the network architecture simple and use a set of data tokens that operate at different temporal resolutions in a hierarchical manner, accounting for the temporally hierarchical nature of videos. By employing a simple two-tower architecture, we are able to initialize our video-language model with pretrained image-text foundation models, thereby boosting the final performance. Second, existing video-language alignment works struggle due to the lack of semantically aligned large-scale training data. To overcome it, we leverage recent LLMs to curate the largest video-language dataset to date with better visual grounding. Furthermore, unlike existing video-text datasets which only contain short clips, our dataset is enriched with video clips of varying durations to aid our temporally hierarchical data to-kens in extracting better representations at varying temporal scales. Overall, empirical results show that our proposed approach surpasses state-of-the-art methods on Multiple retrieval benchmarks, especially on longer videos, and performs competitively on classification benchmarks.

Diffusion-based generative modeling has been achieving state-of-the-art results on various generation tasks. Most diffusion models, however, are limited to a single-generation modeling. Can we generalize diffusion models with the ability of multi-modal generative training for more generalizable modeling? In this paper, we propose a principled way to define a diffusion model by constructing a unified multi-modal diffusion model in a common {\em diffusion space}. We define the forward diffusion process to be driven by an information aggregation from multiple types of task-data, {\it e.g.}, images for a generation task and labels for a classification task. In the reverse process, we enforce information sharing by parameterizing a shared backbone denoising network with additional modality-specific decoder heads. Such a structure can simultaneously learn to generate different types of multi-modal data with a multi-task loss, which is derived from a new multi-modal variational lower bound that generalizes the standard diffusion model. We propose several multi-modal generation settings to verify our framework, including image transition, masked-image training, joint image-label and joint image-representation generative modeling. Extensive experimental results on ImageNet indicate the effectiveness of our framework for various multi-modal generative modeling, which we believe is an important research direction worthy of more future explorations.

Image to image matching has been well studied in the computer vision community. Previous studies mainly focus on training a deep metric learning model matching visual patterns between the query image and gallery images. In this study, we show that pure image-to- image matching suffers from false positives caused by matching to local visual patterns. To alleviate this issue, we propose to leverage recent advances in vision-language pretraining research. Specifically, we introduce additional image-text alignment losses into deep metric learning, which serve as constraints to the image-to-image matching loss. With additional alignments between the text (e.g., product title) and image pairs, the model can learn concepts from both modalities explicitly, which avoids matching low-level visual features. We progressively develop two variants, a 3-tower and a 4-tower model, where the latter takes one more short text query input. Through extensive experiments, we show that this change leads to a substantial improvement to the image to image matching problem. We further leveraged this model for multimodal search, which takes both image and reformulation text queries to improve search quality. Both offline and online experiments show strong improvements on the main metrics. Specifically, we see 4.95% relative improvement on image matching click through rate with the 3-tower model and 1.13% further improvement from the 4-tower model.

To evaluate the performance of a multi-domain goal-oriented Dialogue System (DS), it is important to understand what the users’ goals are for the conversations and whether those goals are successfully achieved. The success rate of goals directly correlates with user satisfaction and perceived usefulness of the DS. In this paper, we propose a novel automatic dialogue evaluation framework that jointly performs two tasks: goal segmentation and goal success prediction. We extend the RoBERTa-IQ model (Gupta et al., 2021) by adding multi-task learning heads for goal segmentation and success prediction. Using an annotated dataset from a commercial DS, we demonstrate that our proposed model reaches an accuracy that is on-par with single-pass human annotation comparing to a three-pass gold annotation benchmark.

Knowledge distillation is typically conducted by training a small model (the student) to mimic a large and cumbersome model (the teacher). The idea is to compress the knowledge from the teacher by using its output probabilities as soft-labels to optimize the student. However, when the teacher is considerably large, there is no guarantee that the internal knowledge of the teacher will be transferred into the student; even if the student closely matches the soft-labels, its internal representations may be considerably different. This internal mismatch can undermine the generalization capabilities originally intended to be transferred from the teacher to the student. In this paper, we propose to distill the internal representations of a large model such as BERT into a simplified version of it. We formulate two ways to distill such representations and various algorithms to conduct the distillation. We experiment with datasets from the GLUE benchmark and consistently show that adding knowledge distillation from internal representations is a more powerful method than only using soft-label distillation.

This paper presents animated pose templates (APTs) for detecting short-term, long-term, and contextual actions from cluttered scenes in videos. Each pose template consists of two components: 1) a shape template with deformable parts represented in an And-node whose appearances are represented by the Histogram of Oriented Gradient (HOG) features, and 2) a motion template specifying the motion of the parts by the Histogram of Optical-Flows (HOF) features. A shape template may have more than one motion template represented by an Or-node. Therefore, each action is defined as a mixture (Or-node) of pose templates in an And-Or tree structure. While this pose template is suitable for detecting short-term action snippets in two to five frames, we extend it in two ways: 1) For long-term actions, we animate the pose templates by adding temporal constraints in a Hidden Markov Model (HMM), and 2) for contextual actions, we treat contextual objects as additional parts of the pose templates and add constraints that encode spatial correlations between parts. To train the model, we manually annotate part locations on several keyframes of each video and cluster them into pose templates using EM. This leaves the unknown parameters for our learning algorithm in two groups: 1) latent variables for the unannotated frames including pose-IDs and part locations, 2) model parameters shared by all training samples such as weights for HOG and HOF features, canonical part locations of each pose, coefficients penalizing pose-transition and part-deformation. To learn these parameters, we introduce a semi-supervised structural SVM algorithm that iterates between two steps: 1) learning (updating) model parameters using labeled data by solving a structural SVM optimization, and 2) imputing missing variables (i.e., detecting actions on unlabeled frames) with parameters learned from the previous step and progressively accepting high-score frames as newly labeled examples. This algorithm belongs to a family of optimization methods known as the Concave-Convex Procedure (CCCP) that converge to a local optimal solution. The inference algorithm consists of two components: 1) Detecting top candidates for the pose templates, and 2) computing the sequence of pose templates. Both are done by dynamic programming or, more precisely, beam search. In experiments, we demonstrate that this method is capable of discovering salient poses of actions as well as interactions with contextual objects. We test our method on several public action data sets and a challenging outdoor contextual action data set collected by ourselves. The results show that our model achieves comparable or better performance compared to state-of-the-art methods.

In this paper, we present a model for learning atomic actions for complex activities classification. A video sequence is first represented by a collection of visual interest points. The model automatically clusters visual words into atomic actions based on their co-occurrence and temporal proximity using an extension of Hierarchical Dirichlet Process (HDP) mixture model. Our approach is robust to noisy interest points caused by various conditions because HDP is a generative model. Based on the atomic actions learned from our model, we use both a Naive Bayesian and a linear SVM classifier for activity classification. We first use a synthetic example to demonstrate the intermediate result, then we apply on the complex Olympic Sport 16-class dataset and show that our model outperforms other state-of-art methods.

In this paper, we learn a reconfigurable template for detecting vehicles and classifying their types. We adopt a popular design for the part based model that has one coarse template covering entire object window and several small high-resolution templates representing parts. The reconfigurable template can learn part configurations that capture the spatial correlation of features for a deformable part based model. The features of templates are Histograms of Gradients (HoG). In order to better describe the actual dimensions and locations of “parts” (i.e. features with strong spatial correlations), we design a dictionary of rectangular primitives of various sizes, aspect-ratios and positions. A configuration is defined as a subset of non-overlapping primitives from this dictionary. To learn the optimal configuration using SVM amounts, we need to find the subset of parts that minimize the regularized hinge loss, which leads to a non-convex optimization problem. We solve this problem by replacing the hinge loss with a negative sigmoid loss that can be approximately decomposed into losses (or negative sigmoid scores) of individual parts. In the experiment, we compare our method empirically with group lasso and a state of the art method [7] and demonstrate that models learned with our method outperform others on two computer vision applications: vehicle localization and vehicle model recognition.

We study the problem of automatically learning event AND-OR grammar from videos of a certain environment, e.g. an office where students conduct daily activities. We propose to learn the event grammar under the information projection and minimum description length principles in a coherent probabilistic framework, without manual supervision about what events happen and when they happen. Firstly a predefined set of unary and binary relations are detected for each video frame: e.g. agent's position, pose and interaction with environment. Then their co-occurrences are clustered into a dictionary of simple and transient atomic actions. Recursively these actions are grouped into longer and complexer events, resulting in a stochastic event grammar. By modeling time constraints of successive events, the learned grammar becomes context-sensitive. We introduce a new dataset of surveillance-style video in office, and present a prototype system for video analysis integrating bottom-up detection, grammatical learning and parsing. On this dataset, the learning algorithm is able to automatically discover important events and construct a stochastic grammar, which can be used to accurately parse newly observed video. The learned grammar can be used as a prior to improve the noisy bottom-up detection of atomic actions. It can also be used to infer semantics of the scene. In general, the event grammar is an efficient way for common knowledge acquisition from video.

This paper considers the problem of detecting actions from cluttered videos. Compared with the classical action recognition problem, this paper aims to estimate not only the scene category of a given video sequence, but also the spatial-temporal locations of the action instances. In recent years, many feature extraction schemes have been designed to describe various aspects of actions. However, due to the difficulty of action detection, e.g., the cluttered background and potential occlusions, a single type of features cannot solve the action detection problems perfectly in cluttered videos. In this paper, we attack the detection problem by combining multiple Spatial-Temporal Interest Point (STIP) features, which detect salient patches in the video domain, and describe these patches by feature of local regions. The difficulty of combining multiple STIP features for action detection is two folds: First, the number of salient patches detected by different STIP methods varies across different salient patches. How to combine such features is not considered by existing fusion methods. Second, the detection in the videos should be efficient, which excludes many slow machine learning algorithms. To handle these two difficulties, we propose a new approach which combines Gaussian Mixture Model with Branch-and-Bound search to efficiently locate the action of interest. We build a new challenging dataset for our action detection task, and our algorithm obtains impressive results. On classical KTH dataset, our method outperforms the state-of-the-art methods.

In this paper, we present an image parsing to text description (I2T) framework that generates text descriptions of image and video content based on image understanding. The proposed I2T framework follows three steps: 1) input images (or video frames) are decomposed into their constituent visual patterns by an image parsing engine, in a spirit similar to parsing sentences in natural language; 2) the image parsing results are converted into semantic representation in the form of Web ontology language (OWL), which enables seamless integration with general knowledge bases; and 3) a text generation engine converts the results from previous steps into semantically meaningful, human readable, and query-able text reports. The centerpiece of the I2T framework is an and-or graph (AoG) visual knowledge representation, which provides a graphical representation serving as prior knowledge for representing diverse visual patterns and provides top-down hypotheses during the image parsing. The AoG embodies vocabularies of visual elements including primitives, parts, objects, scenes as well as a stochastic image grammar that specifies syntactic relations (i.e., compositional) and semantic relations (e.g., categorical, spatial, temporal, and functional) between these visual elements. Therefore, the AoG is a unified model of both categorical and symbolic representations of visual knowledge. The proposed I2T framework has two objectives. First, we use semiautomatic method to parse images from the Internet in order to build an AoG for visual knowledge representation. Our goal is to make the parsing process more and more automatic using the learned AoG model. Second, we use automatic methods to parse image/video in specific domains and generate text reports that are useful for real-world applications. In the case studies at the end of this paper, we demonstrate two automatic I2T systems: a maritime and urban scene video surveillance system and a real-time automatic driving scene understanding system.

In this paper, we present a Deformable Action Template (DAT) model that is learnable from cluttered real-world videos with weak supervisions. In our generative model, an action template is a sequence of image templates each of which consists of a set of shape and motion primitives (Gabor wavelets and optical-flow patches) at selected orientations and locations. These primitives are allowed to slightly perturb their locations and orientations to account for spatial deformations. We use a shared pursuit algorithm to automatically discover a best set of primitives and weights by maximizing the likelihood over one or more aligned training examples. Since it is extremely hard to accurately label human actions from real-world videos, we use a threestep semi-supervised learning procedure. 1) For each human action class, a template is initialized from a labeled (one bounding-box per frame) training video. 2) The template is used to detect actions from other training videos of the same class by a dynamic space-time warping algorithm, which searches a best match between the template and target video in 5D space (x, y, scale, ttemplate and ttarget) using dynamic programming. 3) The template is updated by the shared pursuit algorithm over all aligned videos. The 2nd and 3rd steps iterate several times to arrive at an optimal action template. We tested our algorithm on a cluttered action dataset (the CMU dataset) and achieved favorable performance than [7]. Our classification performance on the KTH dataset is also comparable to state-of-the-arts. 1.

In this paper we present a novel method for parsing aerial images with a hierarchical and contextual model learned in a statistical framework. We learn hierarchies at the scene and object levels to handle the difficult task of representing scene elements at different scales and add contextual constraints to resolve ambiguities in the scene interpretation. This allows the model to rule out inconsistent detections, like cars on trees, and to verify low probability detections based on their local context, such as small cars in parking lots. We also present a two-step algorithm for parsing aerial images that first detects object-level elements like trees and parking lots using color histograms and bag-of-words models, and objects like roofs and roads using compositional boosting, a powerful method for finding image structures. We then activate the top-down scene model to prune false positives from the first stage. We learn this scene model in a minimax entropy framework and show unique samples from our prior model, which capture the layout of scene objects. We present experiments showing that hierarchical and contextual information greatly reduces the number of false positives in our results.