Open SlimeVRX opened 2 years ago
Thiếu sót, mọi người không hiểu công việc của mình hoạt động như thế nào
Phương pháp làm dữ liệu dựa trên máy học
Trainning dựa trên chuỗi hành động AMASS
Học LSTM
Có 5 khối LSTM
So sánh người độ cao khác nhau ảnh hưởng như thế nào đến việc Predict
Các khối trong Transpose model Viết lại LSTM, các nút nhớ, quên
Phải thể hiện rõ vừa ước tính tư thế, vừa ước tính di chuyển trong không gian 3D
Transpose và DIP: So sánh ước tính tư thế
Transpose và DIP: So sánh chuyển động trong không gian 3D
Mục tiêu: Ứng dụng vào game VRChat
Mục tiêu: Thế giới công nghệ luôn thay đổi và phát triển. Những tiến bộ trong công nghệ giúp kết nối giữa con người và con người trở nên dễ dàng hơn, bắt đầu từ gửi mẫu tin nhắn bằng mã Morse đến phát minh ra điện thoại. Hiện nay, nhờ sự phát triển của Metaverse, trong thế giới ảo con người giao tiếp với nhau qua hành động của cơ thể. Ước tính cơ thể người từ cảm biến quán tính phù hợp với xu hướng này. Các ứng dụng AR, VR, game Play to Move yêu cầu hoạt động mọi lúc mọi nơi khiến ước tính cơ thể người từ thị giác máy tính không còn phù hợp.
Tôi muốn chứng minh mạng neural hoạt động với nhiều kích thước cơ thể khác nhau, chiều cao khác nhau 1m6, 1m7, 1m77
Tôi muốn chứng minh, model có thể hoạt đông với cảm biến giá rẻ
Tôi muốn chứng minh, tối ưu tốc độ, độ chính giảm nhưng nằm trong ngưỡng chấp nhận được
Cách tính độ delay?
1 slide so sánh
Bộ dữ liệu SMPL
Lỗi của Transpose model
Lỗi của Transpose Light model
Định tính, định lượng, live demo
317572/ 322999 frame (98.32%) error: 5427 frame
tensor([[13.9808, 6.8008],
[ 7.1220, 3.6887],
[ 4.7241, 2.7607],
[ 5.4321, 3.1413],
VRchat là một game mà người chơi có tương tác với nhau bằng hành
Bộ thực tế ảo thông thường chỉ hỗ trợ theo dõi đầu (HMD) và 2 tay (controller) Một hệ thống hỗ trợ full body tracking trong game VRChat là một thách thức
Thực hiện: Tái tạo lại kết quả của bài báo transpose: