Action Recognition on Input videos:
[ Full Size ]
[ Full Size ]
[ Full Size ]
[ Full Size ]
NOTES:
- The above videos portray the model's atomic action recognition capabilities.
- Active actions are shown with red bounding boxes that are accompanied with the detected action label.
- For the 'Contain' action, the opposite object index is also labelled on corresponding objects.
- These results are preliminary, and are for videos with static camera angles only.
- The input videos were generated by modifying the original CATER Blender code in order to obtain object bounding boxes.
- The bounding boxes are taken from the object annotations (this often leads to occluded objects being labelled in the above samples).
Action Trajectories:
- Each image below compares the actual and predicted Action Trajectory for an input video.
- The X-axis represents the frame index, and the Y-axis denotes the object label.
- Each rectangle represents one action, with it's color representing the action type and it's width denoting the start/end frame.
- For each object instance, the Shaded rectangles (above the object's central line) show the actual actions that the object did.
- The Unshaded rectangles (below the object's central line) show the actions predicted by the model.
- The legend on the top right provides the action-color mapping.
