Zero-Shot Video Question Answer

NExT-QA is a VideoQA benchmark targeting the explanation of video contents. It challenges QA models to reason about the causal and temporal actions and understand the rich object interactions in daily activities, e.g., "why is the boy crying?" and "How does the lady react after the boy fall backward?". It supports both multi-choice and generative open-ended QA tasks. The videos are untrimmed and the questions usually invoke local video contents for answers.

Video-MME stands for Video Multi-Modal Evaluation. It is the first-ever comprehensive evaluation benchmark specifically designed for Multi-modal Large Language Models (MLLMs) in video analysis¹. This benchmark is significant because it addresses the need for a high-quality assessment of MLLMs' performance in processing sequential visual data, which has been less explored compared to their capabilities in static image understanding. The Video-MME benchmark is characterized by its: 1. Diversity in video types, covering 6 primary visual domains with 30 subfields for broad scenario generalizability. 2. Duration in the temporal dimension, including short-, medium-, and long-term videos ranging from 11 seconds to 1 hour, to assess robust contextual dynamics. 3. Breadth in data modalities, integrating multi-modal inputs such as video frames, subtitles, and audios. 4. Quality in annotations, with rigorous manual labeling by expert annotators for precise and reliable model assessment¹. The benchmark includes 900 videos totaling 256 hours, manually selected and annotated, resulting in 2,700 question-answer pairs. It has been used to evaluate various state-of-the-art MLLMs, including the GPT-4 series and Gemini 1.5 Pro, as well as open-source image and video models¹. The findings from Video-MME highlight the need for further improvements in handling longer sequences and multi-modal data, which is crucial for the advancement of MLLMs¹. (1) [2405.21075] Video-MME: The First-Ever Comprehensive Evaluation .... https://arxiv.org/abs/2405.21075. (2) Video-MME. https://video-mme.github.io/home_page.html. (3) Video-MME: Welcome. https://video-mme.github.io/. (4) undefined. https://doi.org/10.48550/arXiv.2405.21075.

The ActivityNet-QA dataset contains 58,000 human-annotated QA pairs on 5,800 videos derived from the popular ActivityNet dataset. The dataset provides a benchmark for testing the performance of VideoQA models on long-term spatio-temporal reasoning. Source: ActivityNet-QA: A Dataset for Understanding Complex Web Videos via Question Answering

The TVQA dataset is a large-scale video dataset for video question answering. It is based on 6 popular TV shows (Friends, The Big Bang Theory, How I Met Your Mother, House M.D., Grey's Anatomy, Castle). It includes 152,545 QA pairs from 21,793 TV show clips. The QA pairs are split into the ratio of 8:1:1 for training, validation, and test sets. The TVQA dataset provides the sequence of video frames extracted at 3 FPS, the corresponding subtitles with the video clips, and the query consisting of a question and four answer candidates. Among the four answer candidates, there is only one correct answer. Source: Two-stream Spatiotemporal Feature for Video QA Task Image Source: https://arxiv.org/abs/1809.01696

MVBench is a comprehensive Multi-modal Video understanding Benchmark. It was introduced to evaluate the comprehension capabilities of Multi-modal Large Language Models (MLLMs), particularly their temporal understanding in dynamic video tasks. MVBench covers 20 challenging video tasks that cannot be effectively solved with a single frame. It introduces a novel static-to-dynamic method to define these temporal-related tasks. By transforming various static tasks into dynamic ones, it enables the systematic generation of video tasks that require a broad spectrum of temporal skills, ranging from perception to cognition.

EgoSchema is very long-form video question-answering dataset, and benchmark to evaluate long video understanding capabilities of modern vision and language systems. Derived from Ego4D, EgoSchema consists of over 5000 human curated multiple choice question answer pairs, spanning over 250 hours of real video data, covering a very broad range of natural human activity and behavior.

The TGIF-QA dataset contains 165K QA pairs for the animated GIFs from the TGIF dataset [Li et al. CVPR 2016]. The question & answer pairs are collected via crowdsourcing with a carefully designed user interface to ensure quality. The dataset can be used to evaluate video-based Visual Question Answering techniques. Source: GitHub Image Source: https://github.com/YunseokJANG/tgif-qa

The MSR-VTT-QA dataset is a benchmark for the task of Visual Question Answering (VQA) on the MSR-VTT (Microsoft Research Video to Text) dataset. The MSR-VTT-QA benchmark is used to evaluate models on their ability to answer questions based on these videos. It's part of the tasks that this dataset is used for, along with Video Retrieval, Video Captioning, Zero-Shot Video Question Answering, Zero-Shot Video Retrieval, and Text-to-Video Generation.

The MSVD-QA dataset is a Video Question Answering (VideoQA) dataset. It is based on the existing Microsoft Research Video Description (MSVD) dataset, which consists of about 120K sentences describing more than 2,000 video snippets. In the MSVD-QA dataset, Question-Answer (QA) pairs are generated from these descriptions. The dataset is mainly used in video captioning experiments but due to its large data size, it is also used for VideoQA. It contains 1970 video clips and approximately 50.5K QA pairs.

TVQA+ contains 310.8K bounding boxes, linking depicted objects to visual concepts in questions and answers. Source: TVQA+: Spatio-Temporal Grounding for Video Question Answering Image Source: https://github.com/jayleicn/TVQAplus

We contribute an IntentQA dataset with diverse intents in daily social activities. We utilize NExT-QA as the source dataset to construct our dataset. NExT-QA dataset is a comprehensive VideoQA dataset with rich natural daily social activities and detailed QA annotations. Originally, the NExT-QA dataset categorizes itself into three types, i.e., Causal, Temporal, Descriptive. We select the inference QA types, i.e., Causal and Temporal, rather than the factoid Descriptive, to build our IntentQA dataset. Particularly, we select both the Causal Why and Causal How subtypes under Causal, and the Temporal Previous and Temporal Next subtypes under Temporal. The Causal Why (CW) QA usually takes the form of ‘Why [action]? For [intent]’, with the key action appearing in the question and the intent in the answer. On the contrary, the Causal How (CH) QA usually takes the form of ‘How [intent]? By [action]’, with the key action appearing in the answer and the intent in the question. The Temporal Previous (TP) QA usually takes the form of ‘What [action A] before [action B]? ’, while the Temporal Next (TN) QA takes the form of ‘What [action B] after [action A]? ’. In the TP&TN QA, the intent is not explicitly expressed in the question nor answer, but is the implicit causal factor linking the two sequential actions.

We study visually grounded VideoQA in response to the emerging trends of utilizing pretraining techniques for video-language understanding. Specifically, by forcing vision-language models (VLMs) to answer questions and simultaneously provide visual evidence, we seek to ascertain the extent to which the predictions of such techniques are genuinely anchored in relevant video content, versus spurious correlations from language or irrelevant visual context. Towards this, we construct NExT-GQA -- an extension of NExT-QA with 10.5K temporal grounding (or location) labels tied to the original QA pairs. With NExT-GQA, we scrutinize a variety of state-of-the-art VLMs. Through post-hoc attention analysis, we find that these models are weak in substantiating the answers despite their strong QA performance. This exposes a severe limitation of these models in making reliable predictions.

How to capture the present knowledge from surrounding situations and perform reasoning accordingly is crucial and challenging for machine intelligence. STAR Benchmark is a novel benchmark for Situated Reasoning, which provides 60K challenging situated questions in four types of tasks, 140K situated hypergraphs, symbolic situation programs, and logic-grounded diagnosis for real-world video situations. (Data Download, STAR Leaderboard)

Neptune is a dataset consisting of challenging question-answer-decoy (QAD) sets for long videos (up to 15 minutes). The goal of this dataset is to test video-language models for a broad range of long video reasoning abilities, which are provided as "question type" labels for each question, for example "video summarization", "temporal ordering", "state changes" and "creator intent" amongst others.

VNBench is a comprehensive benchmark suite for video generative models, which evaluates video generation quality across specific, hierarchical, and disentangled dimensions, each with tailored prompts and evaluation methods. The suite includes 16 dimensions for evaluating Text-to-Video (T2V) models, such as subject consistency, motion smoothness, and overall consistency. VNBench also supports evaluating Image-to-Video (I2V) models and has recently introduced VBench-Long for evaluating longer videos¹. It's designed to align with human perceptions and provide valuable insights for future developments in video generation. # Diversity of “Needle” Types Edit: Using artificially added subtitles as the "needle". These subtitles are embedded in video frames to simulate the scenario of finding specific textual information in a video. Insert: Using images as the "needle". These images are inserted as static segments between video frames to assess the model's ability to recognize and remember static images in a video. Level Classification: Classified into two levels based on image recognizability. The first level uses common objects (e.g., fruit images), while the second level uses more challenging landmark or object images, increasing the task's difficulty. # Diversity of Video “Haystack” Temporal Distribution: The video "haystack" used by VNBench comes from different data sources, with video durations ranging from 10 seconds to 180 seconds. This covers short, medium, and long video lengths to evaluate the model's adaptability to different video lengths. Content Coverage: The video content includes various scenes, ensuring the evaluation's broadness and the diversity of video sources. # Diversity of Queries Retrieval Task: Requires the model to retrieve specific "needles" from videos, assessing the model's fine-grained understanding and information extraction ability. Ordering Task: Requires the model to identify and order the timestamps of all inserted "needles" in the video, assessing the model's understanding of video temporal dynamics and event sequences. Counting Task: Requires the model to count the occurrences of specific objects in the video, including recognizing and tracking repetitive patterns within and across frames. This assesses the model's understanding of spatial and temporal dimensions. (1) Vchitect/VBench: [CVPR2024 Highlight] VBench - GitHub. https://github.com/Vchitect/VBench. (2) VBench: Comprehensive Benchmark Suite for Video Generative Models. https://arxiv.org/abs/2311.17982. (3) Linux中vdbench的安装与使用_vdbench参数详解-CSDN博客. https://blog.csdn.net/SweeNeil/article/details/95338293. (4) Cinebench 2024 Downloads - Maxon. https://www.maxon.net/zh/downloads/cinebench-2024-downloads. (5) undefined. https://doi.org/10.48550/arXiv.2311.17982.

Large multimodal models (LMMs) are processing increasingly longer and richer inputs. Albeit the progress, few public benchmark is available to measure such development. To mitigate this gap, we introduce LongVideoBench, a question-answering benchmark that features video-language interleaved inputs up to an hour long. Our benchmark includes 3,763 varying-length web-collected videos with their subtitles across diverse themes, designed to comprehensively evaluate LMMs on long-term multimodal understanding. To achieve this, we interpret the primary challenge as to accurately retrieve and reason over detailed multimodal information from long inputs. As such, we formulate a novel video question-answering task termed referring reasoning. Specifically, as part of the question, it contains a referring query that references related video contexts, called referred context. The model is then required to reason over relevant video details from the referred context. Following the paradigm of referring reasoning, we curate 6,678 human-annotated multiple-choice questions in 17 fine-grained categories, establishing one of the most comprehensive benchmarks for long-form video understanding. Evaluations suggest that the LongVideoBench presents significant challenges even for the most advanced proprietary models (e.g. GPT-4o, Gemini-1.5-Pro, GPT-4-Turbo), while their open-source counterparts show an even larger performance gap. In addition, our results indicate that model performance on the benchmark improves only when they are capable of processing more frames, positioning LongVideoBench as a valuable benchmark for evaluating future-generation long-context LMMs.

A short clip of video may contain progression of multiple events and an interesting story line. A human needs to capture both the event in every shot and associate them together to understand the story behind it. In this work, we present a new multi-shot video understanding benchmark Shot2Story with detailed shot-level captions and comprehensive video summaries. To facilitate better semantic understanding of videos, we provide captions for both visual signals and human narrations. We design several distinct tasks including single-shot video and narration captioning, multi-shot video summarization, and video retrieval with shot descriptions. Preliminary experiments show some challenges to generate a long and comprehensive video summary. Nevertheless, the generated imperfect summaries can already significantly boost the performance of existing video understanding tasks such as video question-answering, promoting an underexplored setting of video understanding with detailed summaries.

CinePile is a question-answering-based, long-form video understanding dataset. It has been created using advanced large language models (LLMs) with human-in-the-loop pipeline leveraging existing human-generated raw data. It consists of approximately 300,000 training data points and 5,000 test data points.