Download presentation
Presentation is loading. Please wait.
1
Report 7 Brandon Silva
2
Updates Network is training on UAV and producing promising results (512 x 512 input) Multiplied outputs and target by a constant factor before passing through to MSE loss function helps overcome the loss just dropping to zero Same learning rate from paper worked best for training, lower learning rates trains very slowly Gaussian Heatmaps fixed, using method from the paper with help of Waqas Model is not learning from WPAFB Pretty sure ClusterNet network is needed as the labels are too sparse for FoveaNet to train on FoveaNet Examples: Labels I used:
4
Predicted: Target:
5
Predicted: Target:
6
Predicted: Target:
7
Predicted: Target:
8
Predicted: Target:
9
Predicted: Target:
10
Predicted: Target:
11
Next Week Analyze current model and find where a new method could help improved results Start thinking about tracking method for the UAVs
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.