Presentation is loading. Please wait.

Presentation is loading. Please wait.

Appearance Transformer (AT)

Similar presentations


Presentation on theme: "Appearance Transformer (AT)"— Presentation transcript:

1 Appearance Transformer (AT)
Cross-View Action Synthesis Kara Schatz, Dr. Yogesh Rawat, Dr. Mubarak Shah Goal and Motivation Approach Results Synthesize action videos from an unseeen view Input: video sequence from different view single image of desired view Humans can perform this task easily, can machines? Consistency loss between views Keypoint detection Transformation of video and appearance features Using NTU dataset with randomly chosen views Without Appearance Transformer and Keypoints – 400 epochs Input Ground Truth Network Architecture Output View 1 App. View 1 Trans. App. 1 Appearance Transformer Reconstruction Loss (MSE) VGG Input Video View 2 Est. Rep. View 1 Generator Video View 1 Ground Truth Rep. View 1 I3D Est. KP View 1 Core Network Output Video View 2 Rep. View 2 I3D Est. KP View 2 With Appearance Transformation, no Keypoints – 109 epochs Video View 2 Est. Rep. View 2 Generator Input View 2 Reconstruction Loss (MSE) App. View 2 Trans. App. 2 Appearance Transformer VGG Ground Truth Output Core Network Appearance Transformer (AT) Rep. View 1 Input Keypoint Predictor KP View 1 Est. Rep. timestep 1 Est. Rep. timestep 2 Est. Rep. timestep n Consistency Loss (MSE) Consistency Loss (MSE) Ground Truth viewpoint change Trans- former Est. Rep. View 1 Keypoint Predictor Est. KP View 1 Appear-ance AT Cell AT Cell . . . AT Cell Output viewpoint change Trans-former Est. Rep. View 2 Keypoint Predictor Est. KP View 2 Future Work Consistency Loss (MSE) Consistency Loss (MSE) Trans. App. timestep 1 Trans. App. timestep 2 Trans. App. timestep n Improve appearance transformation module Use pixel-weighted loss Use attention Rep. View 2 Keypoint Predictor KP View 2


Download ppt "Appearance Transformer (AT)"

Similar presentations


Ads by Google