This research aims to comprehensively explore building a multimodal foundation model for egocentric video understanding. To achieve this goal, we work on three fronts. First, as there is a lack of QA data for egocentric video understanding, we automatically generate 7M high-quality QA samples for egocentric videos ranging from 30 seconds to one hour long in Ego4D based on human-annotated data. This is one of the largest egocentric QA datasets. Second, we contribute a challenging egocentric QA benchmark with 629 videos and 7,026 questions to evaluate the models' ability in recognizing and memorizing visual details across videos of varying lengths. We introduce a new de-biasing evaluation method to help mitigate the unavoidable language bias present in the models being evaluated. Third, we propose a specialized multimodal architecture featuring a novel ``Memory Pointer Prompting" mechanism. This design includes a global glimpse step to gain an overarching understanding of the entire video and identify key visual information, followed by a fallback step that utilizes the key visual information to generate responses. This enables the model to more effectively comprehend extended video content. With the data, benchmark, and model, we build MM-Ego, an egocentric multimodal LLM that shows powerful performance on egocentric video understanding.

Supervised multi-class learning arises in many application domains such as biology, computer vision, social network analysis, and information retrieval. These applications often involve high-dimensional data, which not only significantly increase the time and space requirement of the underlying algorithms but also degrade their performance due to the curse of dimensionality. Feature selection has been proven effective and efficient for preparing high-dimensional data for many learning tasks. Traditional feature selection algorithms for multi-class data assume the independence of label categories and select features with the capability to distinguish samples from different classes. However, class labels in multi-class data may be correlated and little work exists for exploiting label correlation in multi-class feature selection. In this paper, we investigate label correlation in feature selection for multi-class data. In particular, we provide a principled approach for capturing label correlation and propose an Embedded Supervised Feature Selection (ESFS) framework, which embeds label correlation modeling in supervised feature selection for multi-class data. Experiments on both synthetic data and various types of public benchmark datasets show that the proposed framework effectively captures the multi-class label correlation and significantly outperforms existing state-of-the-art baseline methods.

Video-based coaching systems have seen increasing adoption in various applications including dance, sports, and surgery training. Most existing systems are either passive (for data capture only) or barely active (with limited automated feedback to a trainee). In this paper, we present a video-based skill coaching system for simulation-based surgical training by exploring a newly proposed problem of instructive video retrieval. By introducing attribute learning into video for high-level skill understanding, we aim at providing automated feedback and providing an instructive video, to which the trainees can refer for performance improvement. This is achieved by ensuring the feedback is weakness-specific, skill-superior and content-similar. A suite of techniques was integrated to build the coaching system with these features. In particular, algorithms were developed for action segmentation, video attribute learning, and attribute-based video retrieval. Experiments with realistic surgical videos demonstrate the feasibility of the proposed method and suggest areas for further improvement.

User-adaptive image retrieval/recommendation has drawn a lot of research interests in recent years, owing to fast development of various Web applications where retrieving images is a key enabling task. Existing challenges include the lack of user-adaptive training data, the ambiguity of user query and the real-time interactivity of a system. This paper proposes a hybrid learning strategy that fuses knowledge from both pointwise and pairwise training data into one framework for attribute-based, user-adaptive image retrieval. Under this framework, we develop an online learning algorithm for updating the ranking performance based on user feedback. Furthermore, we derive the framework into a kernel form, allowing easy application of kernel techniques. The proposed approach is evaluated on two image datasets and experimental results show that it achieves obvious performance gains over ranking and zero-shot learning from either type of training data independently. In addition, the online learning algorithm is able to deliver much better performance than batch learning, given the same elapsed running time, or can achieve better performance in much less time.

Relative attributes learning aims to learn ranking functions describing the relative strength of attributes. Most of current learning approaches learn ranking functions for each attribute independently without considering possible intrinsic relatedness among the attributes. For a problem involving multiple attributes, it is reasonable to assume that utilizing such relatedness among the attributes would benefit learning, especially when the number of labeled training pairs are very limited. In this paper, we proposed a relative multi-attribute learning framework that integrates relative attributes into a multi-task learning scheme. The formulation allows us to exploit the advantages of the state-of-the-art regularization-based multi-task learning for improved attribute learning. In particular, using joint feature learning as the case studies, we evaluated our framework with both synthetic data and two real datasets. Experimental results suggest that the proposed framework has clear performance gain in ranking accuracy and zero-shot learning accuracy over existing methods of independent relative attributes learning and multi-task learning.

Video-based systems have been increasingly used in various training tasks in applications like sports, dancing, and surgery. One key task to add automation to such systems is to automatically select reference videos for a given training video of a trainee. In this paper, we formulate a new problem of instructive video retrieval and propose a solution using both attribute learning and learning to rank. The method first evaluates a user's skill attributes by relative attribute learning. Then, the most critical skill attribute in need of improvement is selected and reported to the user. Finally, a hybrid ranking learning to rank method is employed to retrieve instructive videos from a dataset, which serve as reference for the user. Two main technical problems are solved in this method. First, we combine both skill and visual feature to characterize skill superiority and context similarity. Second, we propose a hybrid ranking approach that works with both pair-wise and point-wise labels of the data. The benefit of the proposed method over other heuristic methods is demonstrated by both objective and subjective experiments, using surgical training videos as a case study.

Attribute learning has attracted a lot of interests in recent years for its advantage of being able to model high-level concepts with a compact set of midlevel attributes. Real-world objects often demand multiple attributes for effective modeling. Most existing methods learn attributes independently without explicitly considering their intrinsic relatedness. In this paper, we propose max margin multiattribute learning with low-rank constraint, which learns a set of attributes simultaneously, using only relative ranking of the attributes for the data. By learning all the attributes simultaneously through low-rank constraint, the proposed method is able to capture their intrinsic correlation for improved learning; by requiring only relative ranking, the method avoids restrictive binary labels of attributes that are often assumed by many existing techniques. The proposed method is evaluated on both synthetic data and real visual data including a challenging video data set. Experimental results demonstrate the effectiveness of the proposed method.