Video experiment assignment

[NorthGuard]

1

• Chose 5 objects
• Take photos of objects
• Print 5 most probable labels of each object
• Decide on performance (perhaps compute accuracy)
• Did it perform well? Why? where there difference in the labels performance?

2

• Google ImageNet label-list and check whether all your objects are on the list
• Does some of the labels have few datapoints?

3

• Collect your best-working items
• Work with other group to make a large collection of good items
• Test on all your computer!
• Do they have the same performance?
• Why?

[NorthGuard]

From meeting with Lars:

• Make experiments with a few objects and same background - can it recognise them?
• Try with new background - what happens?
• Make a dataset of a common object (for example a pen) across the entire class. Different pens and different background. Can it learn to recognise pens?

[NorthGuard]

Hey Laura. I forgot to tell you about it, but the above comment is from the meeting with Lars about extensions of the Notebook. Perhaps we can discuss their implementation tomorrow? :D

[laura-rieger]

Regarding the network, I do not think it is possible to train a neural network larger than a linear classifier on a laptop. Currently we have a very simple network implemented, another possible solution would be to run it on the cluster.

[NorthGuard]

From this morning:

• Figure about the process of training models
• Widgetification of some settings
• Make crosshair 224x224 (optional setting)
• Store original s in subdirectory with image augmentation
• Ensure correct format of augmentet images
• Make sure the evaluation video has a crosshair and only uses center of image

[laura-rieger]

All except "Ensure correct format of augmentet images" and figuredone (I don't get the error, so we should check on linux. Could maybe use more widgets but that would basically reduce everything to one big black box.

Regarding figure, I think the one we have right now is not ideal.

[laura-rieger]

fixed augmentation bug and added image

[NorthGuard]

A few remarks before closing exercise: "Image Demo 1"

• Cells 3 and 4 are basically identical. Why do we have two?
• The file-vs-webcam interface is awesome! - I think it should default to webcam though
• In the VideoTakeDashboard we can select a detection algorithm - why? I am looking into some of the fixes today.

"Image Demo 2"

• Found a small bug and some warnings that I will remove
• Looks really sweet :P

[NorthGuard]

We need to be able to lead images back in, so that if something crashed the students do not have to start over.

[NorthGuard]

I have looked a lot into the Image Demo 2 today. The performance of the system is very bad. Also we agreed with Lars to use a pretrained network and only retrain the last layer - why are we not doing this? @laura-rieger @SkafteNicki

[laura-rieger]

Hi, we tried training that and the training takes an undetermined amount of time (at least ten minutes per epoch). The previous results on Nickis laptop were because he is using a GPU (which I do not think we can assume every student has). Therefore I think it is much better to use a simple network.

[NorthGuard]

Okay I see. The performance is only bad when running the live network. Perhaps there is a problem elsewhere? I have looked at preprocessing, but I don't think that is the problem :/

[SkafteNicki]

Just to follow up, we checked the graph of the network in tensorboard and everything seems correct (only the newly added layers are connected to the node that computes gradients). But it is still weird that it takes so long time. Maybe it has something to do with keras, and we are better of using native tensorflow?

[NorthGuard]

Got some help from Laura - problems seems to be fixed now :D