How to add my own images into dest.h5? about synthtext

[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).

[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?

[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

[CatherineYao]

How to get the "seg" "label" and "area" from gPb-UCM (https://github.com/jponttuset/mcg) code that you advised. Thanks a lot. @ankush-me

[FuriousRococo]

Hello Ankush, I've been stuck in creating segmentation infomation of the images and I really need u help! I've managed to run u code in Windows system with Python 3.5 and OpenCV3, but I couldn't find a way to replace gPb-UCM? Is there any method written in python while feasible to run in Windows?

[ankush-me]

I got segmentation information from gPb-UCM / MCG (as pointed above).
You can use scikit-image (I haven't used this myself): http://scikit-image.org/docs/dev/auto_examples/segmentation/plot_segmentations.html

Hello ankush, I am having trouble creating my own data set. I want create a new dest.h5, but I didn't find the method. Can you help me?

[425183525]

Hello Ankush，I have generated a .mat file, how do I turn a .mat file into a .h5 file?

[T43113452]

Sorry.I don't know that depth is translate persepctive.

[ahappycutedog]

Hello Ankush，I have generated a .mat file, how do I turn a .mat file into a .h5 file?

你好 我想问下你生成这个mat文件是在ubuntu系统上吗 没有保错吗 我这边用ubuntu16.04 上的matlab2016 就出现了这个错误;
WARNING: Database pascal2012 (folder JPEGImages) not found in /path/to/pascal2012/
WARNING: Database SBD (folder images) not found in /path/to/SBD/
WARNING: Database COCO (folder images) not found in /path/to/COCO/
-- You can disable this warning in install.m --
-- Successful installation of MCG. Enjoy! --
Warning: The temporary variable im will be cleared at the beginning of each iteration of the parfor loop.
Any value assigned to it before the loop will be lost. If im is used before it is assigned in the parfor loop, a runtime error will occur.
See Parallel for Loops in MATLAB, "Temporary Variables".

In run_ucm (line 22)
Starting parallel pool (parpool) using the 'local' profile ... connected to 12 workers.
1 of 3
err
2 of 3
err
3 of 3
Analyzing and transferring files to the workers ...done.
1 of 3
err
2 of 3
err
3 of 3
Error using im2ucm (line 39)
An UndefinedFunction error was thrown on the workers for 'loadvar'. This might be because the file containing 'loadvar' is not accessible on the workers. Use addAttachedFiles(pool, files) to specify the
required files to be attached. See the documentation for 'parallel.Pool/addAttachedFiles' for more details.

Error in run_ucm (line 22)
parfor i = 1:numel(imname)

Caused by:
Undefined function 'loadvar' for input arguments of type 'char'.

请问这个知道怎么解决吗

[etsegenet2020]

hi ... i use this code for my project but i get error like this
Traceback (most recent call last):
File "/home/molaligntamiru/Documents/SynthText/synthgen.py", line 658, in render_text
txt_render_res = self.place_text(img,place_masks[ireg],
File "/home/molaligntamiru/Documents/SynthText/synthgen.py", line 500, in place_text
render_res = self.text_renderer.render_sample(font,collision_mask)
File "/home/molaligntamiru/Documents/SynthText/text_utils.py", line 366, in render_sample
text_type = sample_weighted(self.p_text)
File "/home/molaligntamiru/Documents/SynthText/text_utils.py", line 22, in sample_weighted
return p_dict[np.random.choice(ps,p=ps)]
File "mtrand.pyx", line 892, in numpy.random.mtrand.RandomState.choice
ValueError: a must be 1-dimensional or an integer

what can i do?

[Lane689]

Hi guys, @crazylyf , @cjnolet

How did you extract depth images from h5 file, or how are you upload depth images into synthtext script?

Here is my code:

import h5py
import numpy as np
import cv2

save_dir = 'C:/Users.../depth_imgs'

with h5py.File('depth.h5', 'r') as hf:
    image_ds = hf
    for imagename in image_ds.keys():
        IMAGE_arr = image_ds[imagename][()]
        cv2.imwrite(f"{save_dir}/{imagename}", IMAGE_arr)
        cv2.waitKey(1000)
        cv2.destroyAllWindows()

But when I run this code it shows me this error:

Then I try print(IMAGE_arr.dtype, IMAGE_arr.shape) and this gives me a results:

float32 (2, 600, 400)
float32 (2, 600, 493)
float32 (2, 600, 284)

The first one "2" here refers to two different depth estimates by two different neural networks (one trained on mostly indoor images, and the other on outdoor images).
Because of that I can't imwrite( ) images into a folder while images from seg.h5 (segmentation images) I successfully saved in a folder.
Opencv's imwrite expects (height, width, channels).