1 of 39

DEEPTIKTOK________ _THREE METHODS _______FOR TRACING

___VIDEO MEMES

ELENA PILIPETS / UNIVERSITY OF SIEGEN

2 of 39

3 of 39

  1. FOCUS ON DEEP/ALT TIKTOK
  2. DISCUSS THREE METHODS FOR WORKING WITH COLLECTIONS OF SOCIAL VIDEO CONTENT
  3. COMBINE TIKTOK METADATA WITH AI-DRIVEN VIDEO & AUDIO ANALYSIS

4 of 39

UNDERSTANDING TIKTOK VIDEOS

1. TikTok videos are short. 2. TikTok is all about vertical video. 3. Many videos tend to be music focused and use original sounds. 4. TikTok has an extensive library of sound templates. 5. Many videos also apply searchable sound and video effects in combination with personalized text/emoji stickers.

-> TikToks are variously networked/searchable/templatable

5 of 39

Engaging logic of TikTok video memes

Platform Metadata

interactive features

(likes, views, shares, comments,

challenges, duets etc.)

+

timestamps, accounts, hashtags, sounds, captions, stickers, effects, etc.

+

moving images/audio content

TikToks are variously networked/searchable/templatable. The main networker is often the sound, not the hashtag.

6 of 39

The usual frame rate is 30 frames per second -> 1 second of video generates 30 images.

Tiktok video length can range from 3 seconds to 3 minutes.

7 of 39

TIKTOK: THE SOCIAL MOVING IMAGE

  1. Finding the right sampling technique for frame extraction and detection of visual elements

On sampling and visualization techniques see e.g., Manovich 2013, on video analysis with AI see e.g., Arnold et al. 2019)

8 of 39

Google Video Intelligence API, for example, extracts and annotates the middle frame of each shot based on its visual content. The method is very powerful when it comes to complex story arcs and multiple shot transitions. Limitation: Web detection is not included.

9 of 39

Many TikToks are filmed in one shot.

15 seconds of this video generate 450 frames. Frame rate = 30 fps.

Google Video AI label detection provides the labels of the individual objects present in a given video. However it does not contain additional web information, which is crucial for understanding the social aspects of video distribution and TikTok virality.

Google Video AI - Video Content Analysis

10 of 39

Google Vision API labels detect and extract information about entities in an image, across a broad group of categories. (...) Web detection detects web references to an image based on content from pages with similar images.” (Google Cloud Vision API + Memespector GUI (Chao 2021) + see e.g., Hagen 2017; Omena et al. 2021; Pilipets 2021; Rogers et al. 2021; Pearce et al. 2022; Burkhardt & Rogers 2022; Tucci 2022; on GV labels see e.g., Geboers & Van de Wiele 2020).

15 sec/15 frames

11 of 39

12 of 39

Extracted video frames and audio tracks can be analyzed with different AI features.

How can we analyze a collection of videos?

001_mp4_frame0001.jpg

001_mp4_frame0002.jpg

001_mp4_frame0003.jpg

002_mp4_frame0001.jpg

002_mp4_frame0002.jpg

002_mp4_frame0003.jpg

001_mp4_audio.flac

002_mp4_audio.flac

13 of 39

Video Extractor (Chao 2023) offers different sampling techniques for extracting frames (stills) from a folder of input video files. It saves extracted images in a new folder, which then can be analyzed with different computer vision techniques. It also extracts audio tracks for speech-to-text recognition.

14 of 39

Speech-to-text converter (Chao 2023) transcribes audio content as captions using Google Cloud AI. It saves recognized text in a spreadsheet file, which then can be merged with TikTok metadata. The main affordance of this tool for TikTok research is that it (at least partially) helps to work around the problem of the “original sound”.

15 of 39

Audio content: TikTok creators often upload their videos with “original sound” instead of using TikTok’s built-in music library. If the video is linked with a song taken from the TikTok library, the song and author name is listed. However, as soon as content involves audio editing, the sound will be listed as “original sound” and provided with a unique numeric ID. —> The challenge of sound recognition/sense making.

Spreadsheet output with TikTok metadata scraped using Zeeschuimer (Peeters 2021)

16 of 39

Networked soundscapes

Resonant sounds and co-hashtags

1.

17 of 39

Networked soundscapes

#deeptiktok

co-hashtags

listed sounds

original sounds

18 of 39

🎵WOAH

AND THE DEEPFRIED

ANIMAL

DANCE

MEME

19 of 39

NAVIGATING THE VERNACULAR STYLE SPACE OF #DEEPTIKTOK BY CHOOSING DIFFERENT PATHS:

SOUNDS, EFFECTS, EMOJIS, HASHTAGS…

+

AUDIOVISUAL CONTENT/TEMPLATES

20 of 39

Platform Metadata

engagement metrics

(likes, reblogs, comments, etc.)

+

hashtags, timestamps, account names,

captions, sounds, stickers, effects, etc.

Computer Vision Data

e.g., Google Vision API Labels

(objects, colors, bodies, actions, etc. )

+

detection of web entities and pages with matching images; speech detection

digital (moving) images as data multiplicities

How to contextualize based on research questions and using different visualization layouts (plot, stack, grid, network, montage, matrix, etc)?

.jpg .png .gif, .mp4, etc.

21 of 39

#deeptiktok | 951 videos | 18199 frames

Deep TikTok is a branch of Alternative or Elite TikTok. While Alt TikTok highlights the independent nature of diverse categories,

DeepTok sticks with a handful of ultra-original, niche topics, and is known for its deepfried aesthetic.

22 of 39

Effect templates

video content and co-hashtags

2.

23 of 39

#dreamcore

24 of 39

🌳

25 of 39

Effect templates

video content and co-hashtags

26 of 39

areyoucomingtothetree?

27 of 39

🌳 TOP 10 WEIRDCORE TIKTOKS PUBLISHED WITH #🌳

28 of 39

Cross-platform riffing

References across the web

3.

29 of 39

30 of 39

31 of 39

32 of 39

33 of 39

34 of 39

35 of 39

36 of 39

37 of 39

38 of 39

TIKTOK METHODOLOGIES><ETHICS?

1. include close-up visualizations of body images only after careful evaluation that content creators had intended for their posts to be widely circulated 2. account for potential bias that come with AI analysis; look at the tonality/stance of post captions 3. consider alternative methods of data visualization/ethical fabrication, involving creative, bricolage-style transfiguration of original data

ethical fabrication (Markham 2012) | situated data analysis (Rettberg 2020) | data feminism (D’Ignazio & Klein 2021) | TikTok Cultures Research Network

39 of 39

OTHER USEFUL LINKS