Animal action recognition on small pets

초록

Action Recognition in human behavior has made significant progress recently, but Animal Action Recognition has many challenges. Due to the difficulty of creating an action-labeled training dataset, it is hard to study action recognition fields in animals, especially those of small kinds such as dogs or cats. We use the latest Dataset of AIHub, which has more than 5,000,000 images of dogs and cats, frames labeled with actions, and 15 key points. In our work, we focus on action classification with time-sequential keypoints and maskings. We extracted 30 frames to train our model and made an End-to-End model using yolact[5] to masking image and mmpose[4] to get animal key points. For the result, we made it possible to classify animal actions with 72%, 79% of accuracy for cats and dogs. We believe our temporal model provides the possibility of training small animals' behavior and understanding their action patterns in various fields such as animal health care. Code is 1 available at https://github.com/ animalact/AAR_Net