--- Now Updating ---

This repository is the implementation of "Audio-visual Action Recognition using Transformer Fusion Network". This code is based on the GitHub repository of the "Swin Transformer" paper. (https://github.com/microsoft/Swin-Transformer)

Dataset preparation

1. UCF-sound(The subset of UCF-101)

   1. Download UCF-101 dataset
   2. The class list provided below includes audio files within video data. Please separate these classes from the UCF-101 dataset. (CliffDiving/Rafting/SoccerPenalty/BabyCrawling/LongJump/Hammering/HandstandWalking/CuttingInKitchen/StillRings/BoxingPunchingBag/PlayingDhol/Surfing/BrushingTeeth/Archery/IceDancing/MoppingFloor/PlayingFlute/BoxingSpeedBag/ParallelBars/UnevenBars/Typing/PlayingCello/TableTennisShot/BasketballDunk/ApplyLipstick/BalanceBeam/PlayingDaf/SumoWrestling/CricketShot/Knitting/FloorGymnastics/Shotput/WritingOnBoard/ShavingBeard/Haircut/BlowingCandles/PlayingSitar/HeadMassage/FrontCrawl/BodyWeightSquats/BandMarching/FrisbeeCatch/FieldHockeyPenalty/HandstandPushups/BlowDryHair/Bowling/WallPushups/CricketBowling/SkyDiving/HammerThrow)

2. Kinetics-sound(The subset of Kinetics-400)

   1. Download Kinetics-400 dataset
   2. The class list provided below includes audio files within video data. Please separate these classes from the Kinetics-400 dataset. (playingtrumpet/stompinggrapes/shovelingsnow/playingclarinet/strummingguitar/blowingnose/playingxylophone/blowingoutcandles/rippingpaper/tapdancing/bowling/laughing/playingbassguitar/playingviolin/playingkeyboard/playingtrombone/tappingpen/dribblingbasketball/playingdrums/choppingwood/singing/playingbagpipes/mowinglawn/playingorgan/playingpiano/shufflingcards/playingguitar/playingaccordion/tickling/playingharmonica/tappingguitar/playingsaxophone)

3. Extract the frames from the video.

4. Extract the WAV file from the video

The case of model

1. IVA : all elements, includnig a single frame, T frames, and audio
2. IV : single frame and T frames without audio
3. IA : single frame and its corresponding audio
4. VA : T frames and its corresponding audio