Dense Video Captioning

The ActivityNet Captions dataset is built on ActivityNet v1.3 which includes 20k YouTube untrimmed videos with 100k caption annotations. The videos are 120 seconds long on average. Most of the videos contain over 3 annotated events with corresponding start/end time and human-written sentences, which contain 13.5 words on average. The number of videos in train/validation/test split is 10024/4926/5044, respectively. Source: Bidirectional Attentive Fusion with Context Gating for Dense Video Captioning Image Source: https://cs.stanford.edu/people/ranjaykrishna/densevid/

YouCook2 is the largest task-oriented, instructional video dataset in the vision community. It contains 2000 long untrimmed videos from 89 cooking recipes; on average, each distinct recipe has 22 videos. The procedure steps for each video are annotated with temporal boundaries and described by imperative English sentences (see the example below). The videos were downloaded from YouTube and are all in the third-person viewpoint. All the videos are unconstrained and can be performed by individual persons at their houses with unfixed cameras. YouCook2 contains rich recipe types and various cooking styles from all over the world. Source: http://youcook2.eecs.umich.edu/ Image Source: https://competitions.codalab.org/competitions/20594

This data set was prepared from 88 open-source YouTube cooking videos. The YouCook dataset contains videos of people cooking various recipes. The videos were downloaded from YouTube and are all in the third-person viewpoint; they represent a significantly more challenging visual problem than existing cooking and kitchen datasets (the background kitchen/scene is different for many and most videos have dynamic camera changes). In addition, frame-by-frame object and action annotations are provided for training data (as well as a number of precomputed low-level features). Finally, each video has a number of human provided natural language descriptions (on average, there are eight different descriptions per video). This dataset has been created to serve as a benchmark in describing complex real-world videos with natural language descriptions. Source: A Thousand Frames in Just a Few Words: Lingual Description of Videos through Latent Topics and Sparse Object Stitching

A new multitask action quality assessment (AQA) dataset, the largest to date, comprising of more than 1600 diving samples; contains detailed annotations for fine-grained action recognition, commentary generation, and estimating the AQA score. Videos from multiple angles provided wherever available. Source: What and How Well You Performed? A Multitask Learning Approach to Action Quality Assessment

The ViTT dataset consists of human produced segment-level annotations for 8,169 videos. Of these, 5,840 videos have been annotated once, and the rest of the videos have been annotated twice or more. A total of 12,461 sets of annotations are released. The videos in the dataset are from the Youtube-8M dataset. An annotation has the following format: { "id": "FmTp", "annotations": [ { "timestamp": 260, "tag": "Opening" }, { "timestamp": 16000, "tag": "Displaying technique" }, { "timestamp": 23990, "tag": "Showing foot positioning" }, { "timestamp": 55530, "tag": "Demonstrating crossover" }, { "timestamp": 114100, "tag": "Closing" } ] } Source: Video Timeline Tags (ViTT)

VidChapters-7M is a dataset of 817K user-chaptered videos including 7M chapters in total. VidChapters-7M is automatically created from videos online in a scalable manner by scraping user-annotated chapters and hence without any additional manual annotation. It is designed for training and evaluating models for video chapter generation with or without ground-truth boundaries, and video chapter grounding, as well as for video-language pretraining.

Despite impressive advancements in video understanding, most efforts remain limited to coarse-grained or visual-only video tasks. However, real-world videos encompass omni-modal information (vision, audio, and speech) with a series of events forming a cohesive storyline. The lack of multi-modal video data with fine-grained event annotations and the high cost of manual labeling are major obstacles to comprehensive omni-modality video perception. To address this gap, we propose an automatic pipeline consisting of high-quality multi-modal video filtering, semantically coherent omni-modal event boundary detection, and cross-modal correlation-aware event captioning. In this way, we present LongVALE, the first-ever Vision-Audio-Language Event understanding benchmark comprising 105K omni-modal events with precise temporal boundaries and detailed relation-aware captions within 8.4K high-quality long videos. Further, we build a baseline that leverages LongVALE to enable video large language models (LLMs) for omni-modality fine-grained temporal video understanding for the first time. Extensive experiments demonstrate the effectiveness and great potential of LongVALE in advancing comprehensive multi-modal video understanding.

Procedural videos show step-by-step demonstrations of tasks like recipe preparation. Understanding such videos is challenging, involving the precise localization of steps and the generation of textual instructions. Manually annotating steps and writing instructions is costly, which limits the size of current datasets and hinders effective learning. Leveraging large but noisy video-transcript datasets for pre-training can boost performance, but demands significant computational resources. Furthermore, transcripts contain irrelevant content and exhibit style variation compared to instructions written by human annotators. To mitigate both issues, we propose a technique, Sieve-&-Swap, to automatically curate a smaller dataset: (i) Sieve filters irrelevant transcripts and (ii) Swap enhances the quality of the text instruction by automatically replacing the transcripts with human-written instructions from a text-only recipe dataset. The curated dataset, three orders of magnitude smaller than current webscale datasets, enables efficient training of large-scale models with competitive performance.

We construct a fine-grained video-text dataset with 12K annotated high-resolution videos (~400k clips). The annotation of this dataset is inspired by the video script. If we want to make a video, we have to first write a script to organize how to shoot the scenes in the videos. To shoot a scene, we need to decide the content, shot type (medium shot, close-up, etc), and how the camera moves (panning, tilting, etc). Therefore, we extend video captioning to video scripting by annotating the videos in the format of video scripts. Different from the previous video-text datasets, we densely annotate the entire videos without discarding any scenes and each scene has a caption with ~145 words. Besides the vision modality, we transcribe the voice-over into text and put it along with the video title to give more background information for annotating the videos. In addition, we propose Vript-Bench, a new benchmark consisting of three challenging video understanding tasks (which are carefully double-checked by humans): - Vript-CAP (Caption): A benchmark with detailed captions rather than short captions. - Vript-RR (Retrieve then Reason): A video reasoning benchmark by first giving a detailed description of the scene as a hint and then asking questions about details in the scene. - Vript-ERO (Event Re-ordering): A benchmark that tests the temporal understanding by offering the descriptions of scenes located in two/four different timelines of the same video, and asks the model to give the right temporal order of the scenes.