Train localization cues for different datasets

[nightrome]

Hi, I was wondering whether you could make the files available to train the localization cues on different datasets. Specifically I assume that to create the deploy.prototxt you:

• Removed Dropout
• Removed the annotations
• Removed the loss
• And the solver.prototxt was identical to this one

It would also be great if you could share the script to create the .pickle files.

[kolesman]

Hi Holger,

Yes, you are right. In particular, I've removed the dropout after the layer 'fc6' and 'fc7' with dropping rate of 0.5.

The 'solver.prototxt' is identical to the you one you've linked, except for one parameter: the "stepsize" is equal to 4000.

Currently I do not have easy-to-run script to generate localization cues. It will require some polishing and adapting my old code to the new naming conventions of this repo.

[sun11]

The loss function and input image resize or crop is the most important. Could you explain what is "multilabel logistic loss" in detail? AFAIK, there is a squared loss which used in https://arxiv.org/pdf/1406.5726.pdf.

[yaoqi-zd]

@kolesman Hi, do you mean, to train the foreground cues, I only need to change the 'stepsize' in solver.prototxt, and add this deploy.prototxt with:

1. 'ImageData' layer to provide images;
2. annotation layer(like in pylayers) to provide image labels;
3. 'SigmoidCrossEntropy' layer to compute the multilabel logistic loss. Since there already exist dropout after 'fc6' and 'fc7' in this deploy.prototxt ?

[yaoqi-zd]

@nightrome Hi Holger, have you tried to reproduce the initial forground cues by yourself? I have treid to train the classification network by myself but my generated forground cues are not as good as the provided

[nightrome]

@yaoqi-zd No sorry, I haven't pursued this any further.

[danea32]

@nightrome Hi Holger, have you tried to reproduce the initial forground cues by yourself? I have treid to train the classification network by myself but my generated forground cues are not as good as the provided

could you please share me the script to create the .pickle files, i want to train it in my own dataset, thank you very much~

[zmll]

@yaoqi-zd could you please share me the script to create the .pickle files, i want to train the model,but i failed. thank you very much~

[yaoqi-zd]

@yaoqi-zd could you please share me the script to create the .pickle files, i want to train the model,but i failed. thank you very much~

Hi, you can refer to the function 'get_localization_cues_sec' in my code which generates the seeds with cam and external saliency maps.

[zmll]

@yaoqi-zd could you please share me the script to create the .pickle files, i want to train the model,but i failed. thank you very much~

Hi, you can refer to the function 'get_localization_cues_sec' in my code which generates the seeds with cam and external saliency maps.

Ok, I will try it,thank you