conzel / paprika Goto Github PK

Before demo

• Run the app on the new hardware.
• Choose some objects, the 4 layers to show and the exact refresh time. #2 #5
• Crop the ImageNet images for the UI. #3

Can do after demo

• Write a script to start the app automatically when the computer is started.
• Ensure that the ML computations (saliency map, prediction, most active layers) are running on the GPU.
• Check for unreasonably high RAM, CPU and GPU usage.
• Refactor the analysis so that ML elements do not get unnecessarily reinitialised.
• Obtain the heatmap in the size of the original camera image so that it does not look pixelated.
• Make sure that the prediction rows never overflow.
• Communicate the speech bubble texts to the museum.

Optional

• Obtain a few consecutive shots from the camera and show the analysis for the image that has the largest score for one class.

Preprocessing and converting to tensor should be part of the analaysis method, whereas cropping and rgb conversion should take place in the ui.
Testing compatibility.

Write a list with the filter visualisations from the last 3 layers that look meaningful. Also add to the list the ImageNet classes that might activate these filters.

Options: maximum score among the subclasses, their sum, some kind of weighted sum.

• Find suitable items, check reduced class list for inspirations (ebay Kleinanzeigen, stores, ...)
• Pick items that are most robustly classified (try different positions, backgrounds, lightings, ...)
• Test with document camera once it has arrived

Possibly animate the arrows so that they light up when there's a new prediction.

• Generate the prediction tensors for all classes, for images of all sizes.
• Select the most similar images in the analysis.
• Show the images in the UI and crop them so that they fit. (See how much space there is for the images. For this, need to know the length of the longest labels.)
• Find out if there are any images that should not be shown, for example, images with the bathtub and bikini labels.