Long video understanding is inherently challenging for vision-language models (VLMs) because of the extensive number of frames. With each video frame typically expanding into tens or hundreds of tokens, the limited context length of large language models (LLMs) forces the VLMs to perceive the frames sparsely and lose temporal information. To address this, we explore extreme video token compression towards *one token per frame* at the final LLM layer. Our key insight is that heuristic-based compression, widely adopted by previous methods, is prone to information loss, and this necessitates supervising LLM layers into *learnable* and *progressive* modules for *token-level compression* (LP-Comp). Such compression enables our VLM to digest 2x-4x more frames with improved performance. To further increase the token efficiency, we investigate *frame-level compression*, which selects the frames most relevant to the queries via the internal attention scores of the LLM layers, named *question-conditioned compression* (QC-Comp). As a notable distinction from previous studies, we mitigate the position bias of LLM attention in long contexts, *i.e.*, the over-concentration on the beginning and end of a sequence, by splitting long videos into short segments and employing local attention. Collectively, our combined *token-level* and *frame-level* leads to an e**x**treme compression model for long video understanding, named **XComp**, achieving a significantly larger compression ratio and enabling denser frame sampling. Our XComp is finetuned from VideoChat-Flash with a data-efficient *supervised compression tuning* stage that only requires 2.5\% of the supervised fine-tuning data, yet boosts the accuracy from 42.9\% to 46.2\% on LVBench and enhances multiple other long video benchmarks.

Understanding scenes in movies is crucial for a variety of applications such as video moderation, search, and recommendation. However, labeling individual scenes is a time-consuming process. In contrast, movie level metadata (e.g., genre, synopsis, etc.) regularly gets produced as part of the film production process, and is therefore significantly more commonly available. In this work, we propose a novel contrastive learning approach that uses movie metadata to learn a general-purpose scene representation. Specifically, we use movie metadata to define a measure of movie similarity, and use it during contrastive learning to limit our search for positive scene-pairs to only the movies that are considered similar to each other. Our learned scene representation consistently outperforms existing state-of-the-art methods on a diverse set of tasks evaluated using multiple benchmark datasets. Notably, our learned representation offers an average improvement of 7.9% on the seven classification tasks and 9.7% improvement on the two regression tasks in LVU dataset. Furthermore, using a newly collected movie dataset, we present comparative results of our scene representation on a set of video moderation tasks to demonstrate its generalizability on previously less explored tasks.

The learning objective of vision-language approach of CLIP [63] does not effectively account for the noisy many-to-many correspondences found in web-harvested image captioning datasets, which contributes to its compute and data inefficiency. To address this challenge, we introduce a novel training framework based on cross-modal contrastive learning that uses progressive self-distillation and soft image-text alignments to more efficiently learn robust representations from noisy data. Our model distills its own knowledge to dynamically generate soft-alignment targets for a subset of images and captions in every minibatch, which are then used to update its parameters. Extensive evaluation across 14 benchmark datasets shows that our method consistently outperforms its CLIP counterpart in multiple settings, including: (a) zero-shot classification, (b) linear probe transfer, and (c) image-text retrieval, without incurring extra computational cost. Analysis using an ImageNet-based robustness test-bed [70] reveals that our method offers better effective robustness to natural distribution shifts compared to both ImageNet-trained models and CLIP itself. Lastly, pretraining with datasets spanning two orders of magnitude in size shows that our improvements over CLIP tend to scale with number of training examples.

Scenes play a crucial role in breaking the storyline of movies and TV episodes into semantically cohesive parts. However, given their complex temporal structure, finding scene boundaries can be a challenging task requiring large amounts of labeled training data. To address this challenge, we present a self-supervised shot contrastive learning approach (ShotCoL) to learn a shot representation that maximizes the similarity between nearby shots compared to randomly selected shots. We show how to apply our learned shot representation for the task of scene boundary detection to offer state-of-the-art performance on the MovieNet [33] dataset while requiring only ~25% of the training labels, using 9× fewer model parameters and offering 7× faster runtime. To assess the effectiveness of ShotCoL on novel applications of scene boundary detection, we take on the problem of finding timestamps in movies and TV episodes where video-ads can be inserted while offering a minimally disruptive viewing experience. To this end, we collected a new dataset called AdCuepoints with 3, 975 movies and TV episodes, 2.2 million shots and 19, 119 minimally disruptive ad cue-point labels. We present a thorough empirical analysis on this dataset demonstrating the effectiveness of ShotCoL for ad cue-points detection.

We propose a novel framework to register sports-fields as they appear in broadcast sports videos. Unlike previous approaches, we particularly address the challenge of field- registration when: (a) there are not enough distinguishable features on the field, and (b) no prior knowledge is available about the camera. To this end, we detect a grid of key- points distributed uniformly on the entire field instead of using only sparse local corners and line intersections, thereby extending the keypoint coverage to the texture-less parts of the field as well. To further improve keypoint based homography estimate, we differentialbly warp and align it with a set of dense field-features defined as normalized distance- map of pixels to their nearest lines and key-regions. We predict the keypoints and dense field-features simultaneously using a multi-task deep network to achieve computational efficiency. To have a comprehensive evaluation, we have compiled a new dataset called SportsFields which is collected from 192 video-clips from 5 different sports covering large environmental and camera variations. We empirically demonstrate that our algorithm not only achieves state of the art field-registration accuracy but also runs in real-time for HD resolution videos using commodity hardware.

In transfer learning, one seeks to transfer related information from source tasks with sufficient data to help with the learning of target task with only limited data. In this paper, we propose a novel Coupled End-to-end Transfer Learning (CETL) framework, which mainly consists of two convolutional neural networks (source and target) that connect to a shared decoder. A novel loss function, the coupled loss, is used for CETL training. From a theoretical perspective, we demonstrate the rationale of the coupled loss by establishing a learning bound for CETL. Moreover, we introduce the generalized Fisher information to improve multi-task optimization in CETL. From a practical aspect, CETL provides a unified and highly flexible solution for various learning tasks such as domain adaption and knowledge distillation. Empirical result shows the superior performance of CETL on cross-domain and cross-task image classification.

Human age is considered an important biometric trait for human identification or search. Recent research shows that the aging features deeply learned from large-scale data lead to significant performance improvement on facial image-based age estimation. However, age-related ordinal information is totally ignored in these approaches. In this paper, we propose a novel convolutional neural network (CNN)-based framework, ranking-CNN, for age estimation. Ranking-CNN contains a set of basic CNNs, each of which is trained with ordinal age labels. Then, their binary outputs are aggregated for the final age prediction. From a theoretical perspective, we obtain an approximation for the final ranking error, show that it is controlled by the maximum error produced among subranking problems, and thus find a new error bound, which provides helpful guidance for the training and analysis of deep rankers. Based on the new error bound, we theoretically give an explicit formula for the learning of ranking-CNN and demonstrate its convergence using the stochastic approximation method. Moreover, we rigorously prove that ranking-CNN, by considering ordinal relation between ages, is more likely to get smaller estimation errors when compared with multiclass classification approaches. Through extensive experiments, we show that ranking-CNN outperforms other state-of-the-art feature extractors and age estimators on benchmark datasets.

Human age is considered an important biometric trait for human identification or search. Recent research shows that the aging features deeply learned from large-scale data lead to significant performance improvement on facial image-based age estimation. However, age-related ordinal information is totally ignored in these approaches. In this paper, we propose a novel Convolutional Neural Network (CNN)-based framework, ranking-CNN, for age estimation. Ranking-CNN contains a series of basic CNNs, each of which is trained with ordinal age labels. Then, their binary outputs are aggregated for the final age prediction. We theoretically obtain a much tighter error bound for ranking-based age estimation. Moreover, we rigorously prove that ranking-CNN is more likely to get smaller estimation errors when compared with multi-class classification approaches. Through extensive experiments, we show that statistically, ranking-CNN significantly outperforms other state-of-the-art age estimation models on benchmark datasets.