Skip to content

Results

Omid edited this page Apr 6, 2018 · 1 revision

Results

UNET

Image Only

it seems to be doing a good starter job. the accuracy went up to 32 on autolab, then it overfitted and went down

Variance and Image

It was the best combination we tried, it boosted the accuracy to 39.8 on autolab with 80 epochs. other parameters :

  • frames per sample (training) : 2
  • frames per sample (testing) : 2
  • batch size : 3

attempts to train for more epochs or with larger samples resulted in overfittiing and testing accuracy went down

Variance and Image and Optical Flow

This network although more expensive to train did not get anywhere better than 36.4 on autolab before adding epochs seemed to not be contributing much to it's accuracy improvement

Variance and Image and Optical Flow and Optical Flow Magnitude

This network although hoped initial to be the best performing never went above 10% IOU on autolab. the highest epoch to which it was trained is : 110

Image and Optical Flow and Optical Flow Magnitude

This network performed very poor in terms of training accuracy in epochs less than 100, thus was disregarded as per limited time.

Variance and Image and Optical Flow Magnitude