How to add my own images into dest.h5?

[junboLiu]

Hello ankush, I am having trouble creating my own data set. I want add images to dest.h5,but I didn't find the method. Can you provide the example?

[junboLiu]

I mean how to get the depth and segmentation information of image?

[ankush-me]

Hi @junboLiu! You can use depth-from-single image ConvNets and popular segmentation methods like gPb-UCM or "Holistic Edge Detection".

[crazylyf]

There are an in-door and an out-door pre-trained model for FCRN depth predictions. Is there much difference between those models, and which one should I choose?

[cjnolet]

I used both for my data sets. I basically separated my images into different folders for indoor and outdoor scenes. I believe the only difference between the two models is that the outdoor model attempts to find the sky and make it the maximum depth.

…

On Thu, Jan 19, 2017 at 8:33 AM, yfliu ***@***.***> wrote: There are an in-door and an out-door pre-trained model for FCRN depth predictions. Is there much difference between those models, and which one should I choose? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#5 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABL1YHSv5I2XMX_tPRa1-yFsvyrw9j_Pks5rT2YngaJpZM4K_9NM> .

[ankush-me]

The indoor/ outdoor models are expected to perform better on images from their respective domains. However, this may not be too critical for synthetic text placement.

[crazylyf]

@cjnolet @ankush-me
I see, thx~

[crazylyf]

Another question: the depth of images in dset.h5 file have two channels, while the depth predicted by FCRN has only one channel. Is there any additional preprocessing? Best!

[ankush-me]

FCRN in our CVPR work does not predict depth.

The two channels correspond to predictions by the depth-prediction models trained on outdoor and indoor datasets.

[oneOfThePeople]

how you deal with the problem that each image have different size and the depth prediction and the segmentation network required fix size of input image?

[ankush-me]

As far as I recall, there is a limit of 600px for the depth prediction network, (and no size limit at least for the gPb-UCM segmentation method). So I rescale down the image if it is bigger than 600 px; and then bring all the three things -- RGB, depth, segmentation -- to the same size (check here).

[oneOfThePeople]

i see you resize to the depth image and not to the original one . why?
thank alot for your patience

[crazylyf]

Hi @ankush-me
What does the segmentation variables seg, area, and label mean? Does the seg and area refer to the superpixels of proposals and area of the bounding boxes of gPb-UCM? what is label?
Thanks a lot!

[ankush-me]

@oneOfThePeople : I did not want to interpolate the depth image.

@crazylyf : "Seg" is a region labeling, where pixels in each segment have the same label. "Label" is just a vector of unique values in "seg". "Area" is the number of pixels in the each segment, where, the i-th "area" corresponds to the segment with the i-th "label".

[oneOfThePeople]

ok thank alot