AI Animation Team Clip1 Clip2 Clip4 Clip3 Clip5 Clip6 Clip7 - - PowerPoint PPT Presentation

ai animation team clip1 clip2 clip4 clip3 clip5 clip6
SMART_READER_LITE
LIVE PREVIEW

AI Animation Team Clip1 Clip2 Clip4 Clip3 Clip5 Clip6 Clip7 - - PowerPoint PPT Presentation

AI Animation Team Clip1 Clip2 Clip4 Clip3 Clip5 Clip6 Clip7 Clip8 clip9 Automatically Music Dance Animation Neural Music Animation Network Noise In Output Lack of Paired Data Paired match 3D Skeleton Animation Music


slide-1
SLIDE 1

AI Animation Team

slide-2
SLIDE 2
slide-3
SLIDE 3
slide-4
SLIDE 4

Clip1 Clip2 Clip5 Clip3 Clip6 Clip8 clip9 Clip4 Clip7

slide-5
SLIDE 5

Automatically

Music Dance Animation

slide-6
SLIDE 6
slide-7
SLIDE 7

Music Animation

Noise In Output Lack of Paired Data Neural Network

slide-8
SLIDE 8

Choreography Video

huge number of them available match

  • 3D Skeleton Animation
  • Music

Paired

  • 2D Skeleton Animation
  • Music

Paired

slide-9
SLIDE 9

Music Animation

Neural Network Match

2D Animation Feature Pre-recorded 3D Skeleton Animation (3D)

Limited

Break into smaller pieces then reconstruct

“Unlimited”

slide-10
SLIDE 10
  • Position (Pose)
  • Velocity
  • Acceleration
  • Motion Beats

How do we match 2D animation with 3D animation?

Velocity Motion Beats

slide-11
SLIDE 11

Melody Tabs

16 Beats One Cycle

Choreography Labanotation Phase

0.25 Phase 0.5 Phase 0.75 Phase 1.0 Phase

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

Phase

Beat Counts

1.25 Phase 1.5 Phase

slide-12
SLIDE 12

# 7 # 7 # 6 # 6

Segmented Dance Clips

# 1 # 2 # 3 # 4 # 5 # 2 # 4

  • Cut dance cycle into smaller sections

base on motion beats

  • Reconstruct dance cycle from these

smaller sections based on phase

slide-13
SLIDE 13

Music Animation

Neural Network Phase-based Match

2D Animation Feature Pre-recorded 3D Skeleton Animation (3D)

slide-14
SLIDE 14

Audio Track Phase Based Dance Matching Animation Warping Audio-Motion Interpreter Network Dance Animation Recorded Animation Motion Curve Analysis Dance Clip Segments Dance Motion Feature Dance Motion Feature Matched Dance Sequence Dance Video Train

Dance Matching

System Overview

1 2 3

slide-15
SLIDE 15

3D Dance Sequence

Audio Track Video Track Choreography Video Phase Based Dance Matching Audio Features Dance Motion Features Matched Sequence

Human Pose Detect Clip Segments Clip Segments

… … 1 hour mocap dataset

Fine Grained Matching

Animation Warping Audio-Motion Interpreter Network … … Choreography Video Playlist

Video Clip Motion Feature Audio Track

train

Phase

System Overview

Intelligent Choreography Pipeline Structure

1 2 3

slide-16
SLIDE 16

Audio-Motion Interpreter Network

Database Structure

YouTube – Choreography Playlists

~ 200 videos 3-6 minutes / video

  • Professional Choreography Studios
  • K-pop/Urban/Chinese Traditional
  • Varying bpm 120-200

Multiple Dancers

Challenges

Moving Camera Sliced Scenes

2D Pose Extraction 16 hours 36 min 500 hours

slide-17
SLIDE 17

Audio-Motion Interpreter Network

Automatic Pipeline And the Model Choice

Video Filter Slice and Clean Dance Video Music Beat Info Audio Stream Video Steam 2D Pose Extraction Librosa Aubio Mel-spec CQT Beat Speed Info Mapping Network Conv1D/LSTM+Conv2 D Mixture Density Network ConvNets LSTM Mixture Density Networks Phase-based

  • Dancer-focused
  • Stable Camera
  • Continuous
slide-18
SLIDE 18

Audio-Motion Interpreter Network

Problem Definition Modelling The Mapping From Audio Feature to Motion Features

Audio Track

Audio Features

Dance Motion Features Audio-Motion Interpreter Network Phase Based Dance Matching

Mel-spec

Zero-Padding

Temporal and Frequency Different Length based on bpm Fixed Length

slide-19
SLIDE 19

Audio-Motion Interpreter Network

Problem Definition Modelling The Mapping From Audio Feature to Motion Features

Audio Track Audio Features Dance Motion Features Audio-Motion Interpreter Network Phase Based Dance Matching

time type delta angle max vel duration

Mel-spec

Dance Motion Features For each Joint One Motion Curve 8 Major Joints Shoulder/Elbow/Keen/Ankle

slide-20
SLIDE 20

Audio-Motion Interpreter Network

Problem Definition Modelling The Mapping From Audio Feature to Motion Features

slide-21
SLIDE 21

Audio-Motion Interpreter Network

Problem Definition Modelling The Mapping From Audio Feature to Motion Features

Audio Track Audio Features Dance Motion Features Audio-Motion Interpreter Network Phase Based Dance Matching Mel-spec Training Time 6 hours 40mins 84hours 40mins

slide-22
SLIDE 22

Audio-Motion Interpreter Network

Problem Definition Modelling The Mapping From Audio Feature to Motion Features

Audio Track Audio Features Dance Motion Features Audio-Motion Interpreter Network Phase Based Dance Matching Mel-spec Inference Time for 16-beat 18 ms 1500 ms

slide-23
SLIDE 23

Music Waveform Fragments Mel Spectrograms ConvNets Dance Motion Features New Audio Track Dance Motion Features 3D Dance Sequence Audio-Motion Interpreter Fine Grained Dance Matching Waveform Fragments Audio Track Video Track Online Choreography Video Playlist 2D Motion Motion Curve Analysis Module Training

Audio-Motion Interpreter Network

Network Structure and Training procedure

ConvNets

slide-24
SLIDE 24

Beats Distribution Beat Max Velocity Beats Delta Distance Thigh Calf Upper-arm Fore-arm Local Angle Speed Curves Global Guideline Curves Local Feature Curves Video Frames Detected 2D Motions

Lower Body Rhythm Upper Body Rhythm Beat Breakpoints

3D Dance Animation Clips 3D Dance

Clip Retrieval

Beat Strength

Motion Curve Analysis

Phase-based Dance Matching

2.1 Motion Curve Analysis and Retrieval Module

slide-25
SLIDE 25

Melody Tabs

16 Beats One Phase

Choreography Labanotation Phase

0.25 Phase 0.5 Phase 0.75 Phase 1.0 Phase

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

Phase

Beat Counts

1.25 Phase 1.5 Phase

Phase-based Dance Matching

2.2 Choreography Phase Circle

slide-26
SLIDE 26

Inside Phase Breakpoints Distribution

Automatic Dataset Preprocess Segmented Dance Clips

# 1 # 2 # 3 # 4 # 5 # 6 # 7 # 8 Motion Curve Analysis # 1 # 2 # 3 # 4 # 5 # 6 # 7 # 8

Phase-based Clips 256 frames

Fine- grained Slicing Mocap Animation Clip

Phase-based Dance Matching

2.3 Phase-based Fine-grained Data Preparation Phase-based Slicing

Breakpoint #1 Breakpoint #2

slide-27
SLIDE 27

Segmented Dance Clips

2/16 Bests Dance Phase Breakpoint #1 6/16 Bests Dance Phase Breakpoint #2 12/16 Bests

Breakpoint #3

Dance Phase Breakpoint #0 Clip # 4 Phase # 2 Clip # 2 Segment # 1 Clip # 7 Phase # 3 Clip # 8 Phase # 2 Matched Phase Potential Candidates # 1 # 2 # 3 # 4 # 5 # 6 # 7 # 8 # 2 # 4 # 7 # 8

Phase-based Dance Matching

2.4 Phase-based Fine-grained Motion Matching

Dance Motion Features Audio-Motion Interpreter Network Guided Match

Orientation Pose Similarity Velocity Vector Angel Velocity

slide-28
SLIDE 28

Original Animation Beat Sequence Music Beat Sequence Adjusted Animation Beat Sequence

1.0x Original Playback Speed 1.0x Adjusted Playback Speed

Time Warp Mechanism

Align Animation Beats with Music Beats

Animation Warping

Enhance the Emotional Impact

slide-29
SLIDE 29

Adjusted Animation Beat Sequence

Time Warp Mechanism

Animation Warping

Enhance the Emotional Impact

Original Animation Beat Sequence Music Beat Sequence

1.0x Original Playback Speed

Matching Beats According to Strength Align Animation Beats with Music Beats

slide-30
SLIDE 30

Animation Warping

Enhance the Emotional Impact

Time Warping Curve

slide-31
SLIDE 31

Animation Warping

Enhance the Emotional Impact

Time Warping Curve Enhanced Beats

slide-32
SLIDE 32

Animation Warping

Enhance the Emotional Impact

Before Beats Aligned

slide-33
SLIDE 33

Huge Pipeline Efficiency Boost

From GPU Computation

Audio Track Phase Based Dance Matching Animation Warping Audio-Motion Interpreter Network Dance Animation Dance Motion Feature Matched Dance Sequence Dance Video 2D Pose Extraction For Training

1 2

Training Time Phase 0 16 hour Phase 1 6 hour Phase 2 2 hour Phase 0 500 hour Phase 1 84 hour Phase 2 20 hour

slide-34
SLIDE 34

Music > {neural network}>{matching}> Animation

System Performance on UGC platform

slide-35
SLIDE 35

Dance Matching

Online Feedback

10000000+ request UGC Widely Spread

in Multiple Video Platform

slide-36
SLIDE 36

Dance Matching

Online UGC on Bilibiili

slide-37
SLIDE 37

This System is Currently

Powering