Week 5. Testing with models from generated dataset

1 minute read

First tests

So as we thought, the results are not the ones we were expecting and the Neural Network is not predicting the correct position of the ball because the data extract from the videos is quiet similar but not identical at all.

Cap0
Cap1

This happend because how we recorded the videos and made the ball went from the left to the right, the route and speed are not the same in our automated generated datasets. Both data has to be almost the same. It is difficult to achieve the technique becase the DOF in real life is more complicated to follow comparing with a script that match the caracteristics we want to.

In order to do so, we propuse to create new datasets, with a bit changes in the way the route thought pixels is made.

How:

  • Before we only had an array with 20 values of [x,y] -> and 1 value to predict within the gap (30)

  • Now we pretend to have 20 arrays of 20 values meaning we need from array [0-19] to array [20-39] to predict the next 30 frame gap: from frame 50 to 70 are predicted

The real values of the objects now are: (Example)

Inputs

[[[35 0] [35 1] [34 1] [34 2] [34 3] [34 4] [34 4] [33 5] [33 6] [33 8] [33 9] [33 11] [33 12] [33 14] [33 16] [33 18] [32 19] [32 21] [32 22] [32 24] [32 25] [32 27] [32 28] [32 30] [32 32] [32 33] [31 35] [32 36] [31 38] [31 39] [31 41] [31 42] [31 44] [31 45] [31 47] [31 48] [31 49] [31 51] [31 52]]]

And the values to predict are:

Outputs

[[30 68] [30 69] [30 71] [30 72] [30 73] [30 75] [30 76] [30 77] [30 79] [30 80] [30 82] [30 83] [30 84] [30 86] [30 87] [30 88] [30 89] [30 91] [30 92] [30 93]]

You May Also Enjoy