Skip to the content.

GitHub Platform Dependencies GitHub repo size

0. Table of Contents

1. Motivation

The seemingly easy-to-mark visually redundant clips (e.g., intros, outros, recaps, and commercial breaks) have long been existent in correlated videos (e.g., a series of TV episodes, shows, and documentaries). Although mainstreaming video content providers (VCPs) such as Netflix, Amazon Prime Video, HBO Max have provided corresponding functionalities for helping users skip them, they incur poor user experiences in practice. We list the specific undesired symptoms as follows.

2. Solution


SkipStreaming is a high-performance and accurate visual redundancy detection system for correlated videos.

SkipStreaming is built based on the novel perspective of “scenes”, which are the basic story units that compose a video. It extracts scene information via our specially-designed “audio-guided scene sketch” methodology, and selectively compares a small portion of video frames to quickly detect visual redundancy.

This repository hosts the code of SkipStreaming and the video dataset involved in our study.

3. Code

The key component of SkipStreaming is a stand-alone server module implemented with a total of 4K and 300 lines of C/C++ and Python code, which can be seamlessly integrated into mainstream web servers with minimal configurations.

The source code of SkipStreaming is released at


SkipStreaming can be built and run on either Windows or Linux platform. Here we provide the guide on how to build and run it on Windows.

Build SkipStreaming with Visual Studio

  1. Just open SkipStreaming.sln using Visual Studio 2022
  2. Right click the SkipStreaming project in Visual Studio and enter the property window in the popup menu
  3. Switch the platform to Win32 in the configuration
  4. Click the Debug tab and add PATH=%PATH%;$(ProjectDir).\dll to the environment
  5. Switch to x86 in the ribbon column
  6. Build and Execute through Ctrl+F5


Suppose that we have two videos to be compared, one is reference video and the other is query video.

First, run to build scene tree for the comparing videos and output the scene tree information to text files.

python -i <video directory>

Then, run SkipStreaming for each pair of videos and get the results in the output text files, where each line represents the beginning and ending frame of a detected redundant clip.

./SkipStreaming.exe \
-r <path to reference video> \
-x <path to multi-index hash table> \
-q <path to query video> \
-a <path to scene tree> \
-g <path to ground truth file> \
-o <path to output markers>

4. Dataset

We provide the information (in particular the marker information) about our dataset videos collected from major VCPs.

The dataset is available at

VCP Directory Marker Collection Methods
Amazon Prime Video amazon_res We capture all video series URLs in different genres of Amazon Prime Video by selecting the URLs from .av-hover-wrapper elements. Afterwards, we navigate to each video series URL and extract the asin property for each episode, which is then fed into Amazon Prime Video’s API ( to obtain the marker information. The intro/recap/outro timestamps are in the transitionTimecodes of the responded json data.
Disney+ disney_content We directly invoke Disney+’s StandardCollection API to obtain the genre information. For each genre, we query its video series information through the CuratedSet API. Based on this, we invoke its DmcSeriesBundle API to query different seasons of each video series as well as the DmcEpisodes API to obtain video series episodes’ information. Finally, we get the episode’s meta information through the DmcVideo API, and the marker information is inside the milestones fields of the responded json data.
HBO Max hbo_content We directly invoke express-content APIs to get the unique ID of each episode for each video series. Then, we invoke the API to obtain episodes’ meta information. The marker information resides in the annotations fields of the responded json data.
Hulu hulu_play We first navigate to the Hulu’s home page to obtain the genre information of Hulu’s library. For each available genre, we retrieve video series information using Hulu’s view_hubs API. Then, we get the season information for each video series through the series API. Finally, we get the meta information of each episode in the video series through Hulu’s playlist API. The marker information resides in the markers fields of the responded json data.
iQIYI iqiyi_dash We obtain the playlist of all available video series from iQIYI’s home page. Then, we navigate to the playback page of each video series, and intercept its network requests to extract the episode information from its dash API. The marker information is in the responded bt and et fields.
Netflix netflix_meta We first get the URLs of Netflix’s all video genres. Then, we navigate to the genre pages to retrieve all video series URLs under each genre. Such URLs contain the unique ID for each video series, based on which we obtain video meta information from Netflix’s metadata API. The responded data contain the maker information of each episode in the skipMarkers field.
Tencent Video tencent_video All video series are listed on Tencent Video’s video list page. We obtain video series URLs from the list page, and visit each URL to get the unique IDs for each episode of the video series. These IDs are then passed into its fcgi-bin/data?idlist API to get the meta information. The marker information resides in the head_time and tail_time fields of the responded data.
Youku youku_appinfo All video series are also listed on Youku’s video list page, and we also obtain all video series URLs. For each video series, we navigate to its playback page, and visit each episode. During this process, we monitor pages’ network requests through Puppeteer pages’ on response API, and get the returned data of Youku’s API, where we can find the meta information of the video episode. The maker information is in the responded point field.