compphoto / boostingmonoculardepth Goto Github PK

Is there is particular reason to use Midas 2 over Midas 3?

While I was trying to get some depth estimation of some of my own images using the Boostmonoculardepth.ipynb, I ran into the following error:

It seems like the link https://cloudstor.aarnet.edu.au/plus/s/lTIJF4vrvHCAI31/download hosting the LeRes weights is down. Is there another channel for me to get LeRes weights?

Not working in my local

Hi, it seems like this error is occuring often when I try to use the Colab Notebook lately. If I reset the runtime and start over, sometimes it works, and sometimes not. It happens with both MiDaS and LeReS, and I don't think it depends on what file(s) or filetypes are in the inputs folder. However, it does appear to work after this error if I create a new folder for the input images, move the images into that folder, and update the directory in the last code cell to match. But how can this be avoided? Please advise—thank you!

I see this error was mentioned before and a bug was fixed—maybe it's the same issue again? A user also mentioned a .ipynb_checkpoint file in the inputs file, but when I try to run a command to remove it, it's not found. But the fact that using a different folder works suggest maybe it's a related problem?

Traceback (most recent call last): File "run.py", line 580, in <module> run(dataset_, option_) File "run.py", line 107, in run for image_ind, images in enumerate(dataset): File "/content/BoostingMonocularDepth/utils.py", line 215, in __getitem__ return Images(self.rgb_image_dir, self.files, index) File "/content/BoostingMonocularDepth/utils.py", line 148, in __init__ self.rgb_image = read_image(os.path.join(self.root_dir, name)) File "/content/BoostingMonocularDepth/utils.py", line 27, in read_image if img.ndim == 2: AttributeError: 'NoneType' object has no attribute 'ndim'

Thank you for your contributions and for your amazing research.
Sharing the evaluation code was immensely helpful.

I'd like to ask however why do you NaN the values as following?
(line 23 on evaluate.m ) gt_disp(gt_disp==0)=nan;
Wouldn't this cause ending up with nan on future computations? (such as applying immse with an arbitrary matrix)

I was wondering if implementing that as follows would help restore the information or rather disturb the image statistics during conversions:
gt_disp(gt_disp==0) = 1e-6 // or eps

Thank you for your support!

Which one of MiDas, DPT and LeRes is better for monocular depth estimation and takes the shortest time? @miangoleh

can't download the weight file

I've been trying to get it to run on my M1 machine. Remove CUDA dependencies wherever I found them. But now I'm stuck here:

initialize network with normal
loading the model from ./pix2pix/checkpoints/mergemodel/latest_net_G.pth
start processing
processing image 0 : sample1
(2964, 3820, 3)
	 wholeImage being processed in : 2912
 	 	 DEBUG| GPU THRESHOLD REACHED 2912 ---> 1568
Traceback (most recent call last):
  File "run.py", line 581, in <module>
    run(dataset_, option_)
  File "run.py", line 125, in run
    whole_estimate = doubleestimate(img, option.net_receptive_field_size, whole_image_optimal_size,
  File "run.py", line 401, in doubleestimate
    pix2pixmodel.test()
  File "/Users/jan/Documents/ML/BoostingMonocularDepth/pix2pix/models/base_model.py", line 106, in test
    self.forward()
  File "/Users/jan/Documents/ML/BoostingMonocularDepth/pix2pix/models/pix2pix4depth_model.py", line 116, in forward
    self.fake_B = self.netG(self.real_A)  # G(A)
  File "/opt/homebrew/Caskroom/miniforge/base/envs/sketch-simplyfication/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1190, in _call_impl
    return forward_call(*input, **kwargs)
  File "/Users/jan/Documents/ML/BoostingMonocularDepth/pix2pix/models/networks.py", line 473, in forward
    return self.model(input)
  File "/opt/homebrew/Caskroom/miniforge/base/envs/sketch-simplyfication/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1190, in _call_impl
    return forward_call(*input, **kwargs)
  File "/Users/jan/Documents/ML/BoostingMonocularDepth/pix2pix/models/networks.py", line 541, in forward
    return self.model(x)
  File "/opt/homebrew/Caskroom/miniforge/base/envs/sketch-simplyfication/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1190, in _call_impl
    return forward_call(*input, **kwargs)
  File "/opt/homebrew/Caskroom/miniforge/base/envs/sketch-simplyfication/lib/python3.8/site-packages/torch/nn/modules/container.py", line 204, in forward
    input = module(input)
  File "/opt/homebrew/Caskroom/miniforge/base/envs/sketch-simplyfication/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1190, in _call_impl
    return forward_call(*input, **kwargs)
  File "/opt/homebrew/Caskroom/miniforge/base/envs/sketch-simplyfication/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 463, in forward
    return self._conv_forward(input, self.weight, self.bias)
  File "/opt/homebrew/Caskroom/miniforge/base/envs/sketch-simplyfication/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 459, in _conv_forward
    return F.conv2d(input, weight, bias, self.stride,
RuntimeError: Input type (MPSFloatType) and weight type (torch.FloatTensor) should be the same

I don't know where to put a .to("mps") or whatever needs to be done.
Anyone else got it to work?

GPU_threshold = 1600 - 32 # Limit for the GPU (NVIDIA RTX 2080), can be adjusted

how to find this out for a apecific gpu , does it imapact performance ??

is it the numbers of cuda cores of a gpu ??

it's possible to extract an array with numerical depth values (PFDM Midas files) ?
i'm using the python notebook on Colab but in output i have only images

thanks
Luca

Can anyone guide me to solve the following error:

I'm trying to run the "test.py" after training by using this command: "python ./pix2pix/test.py --dataroot DATASETDIR --name mergemodeleval --model pix2pix4depth --no_flip --no_dropout"

but I get this error:

creating web directory /content/drive/MyDrive/BMD/net/BoostingMonocularDepth/pix2pix/checkpoints/mergemodeleval/test_latest
OrderedDict():
OrderedDict()
visual_ret[name]:
tensor([[[[-0.3939, -0.5247, -0.5284, ..., -0.5045, -0.5042, -0.4400],
[-0.5320, -0.5526, -0.5533, ..., -0.5138, -0.5113, -0.5006],
[-0.5219, -0.5484, -0.5297, ..., -0.5128, -0.5091, -0.4998],
...,
[ 0.3818, 0.3915, 0.3934, ..., -0.4403, -0.4513, -0.4354],
[ 0.3901, 0.3910, 0.3884, ..., -0.4387, -0.4599, -0.4186],
[ 0.3347, 0.3884, 0.3821, ..., -0.4316, -0.4241, -0.2628]]]],
device='cuda:0')
visual_ret:
OrderedDict([('fake_B', tensor([[[[-0.3939, -0.5247, -0.5284, ..., -0.5045, -0.5042, -0.4400],
[-0.5320, -0.5526, -0.5533, ..., -0.5138, -0.5113, -0.5006],
[-0.5219, -0.5484, -0.5297, ..., -0.5128, -0.5091, -0.4998],
...,
[ 0.3818, 0.3915, 0.3934, ..., -0.4403, -0.4513, -0.4354],
[ 0.3901, 0.3910, 0.3884, ..., -0.4387, -0.4599, -0.4186],
[ 0.3347, 0.3884, 0.3821, ..., -0.4316, -0.4241, -0.2628]]]],
device='cuda:0'))])
processing (0000)-th image... 00380_colors_&0_0&.png
Traceback (most recent call last):
File "./pix2pix/test.py", line 71, in
inner = visuals['inner']
KeyError: 'inner'

I have a doubt about the evaluation of prediction results on the KITTI dataset. I tried to get the depth estimation of a particular image from the KITTI dataset and observe the result using a simple python code as shown below. The prediction is done using depthNet = 2 (The LeReS model)

depth_gt = Image.open(path)
depth_gt = np.asarray(depth_gt, dtype=np.float32)

I printed the result and observe that the "pixel" value of the image is of range [0, 65535], and the ground truth returns a range of [0, 255]. I mainly want to know the ORD error of the image, may I know if there is any image transformation before I do the evaluation? Or I could provide the directory path into
estimation_path = '';
gt_depth_path = '';
from evaluatedataset.m and get the ORD error directly.

To be more specific the prediction depth map obtained from boosting monocular technique, followed by the ground truth and the original RGB are below. Should I rescale the range of the depth map? or I should perform normalization on both of the images before conducting the evaluation.

Edit: I notice the difference in the "pixel" value is due to a different data type used to represent the image. Some images use float32 to represent their image and the values are in mm while others could be represented by uint8. My question now is how do I use the evaluation code in this case? I am currently doing testing on the DIODE dataset by scaling the ground truth & prediction depth map into uint8 representation before evaluation. Would really like to know if I am doing this correctly.

I can't download the merge net model from the link of your site.

Hey,
I understand that the Local Boosting method is done as follows:
-Upsample image if needed
-choose patch size to be receptive field size where 1/3 receptive field size interleave with other patches
-increase the resolution of patches to have enough context
-discard smooth patches
-sort patches from biggest to smallest

• double estimation on each patch:
  High res is fixed at 2*field size
  Low res is receptive field size
  Then the two patches are resized to 1024x1024 (inference size) not sure about this process TBH
• merge patch with corresponding patch in updated image
  The two are scaled to the same size
  Pix2pixsize * pix2pixsize
• some trick with polynomials is done since estimations are relative
• apply gaussian mask to make merging look seamless

(1) So, in order to merge two images we resize them to similar fixed sizes?
(2) Are the low and high resolutions for patches correct?
(3) Double-estimation is applied to each patch separately and then with the patch in the base estimate, right?
Thanks in advance!

Is it possible to run this with cpu only? Unfortunately I keep getting Cuda out of memory. :(

Discovered in #5

$ codespell --count --ignore-words-list="pres"

17
./BoostingMonocularDepth/run.py:325: patchs ==> patches, paths
./BoostingMonocularDepth/utils.py:34: Guassian ==> Gaussian
./BoostingMonocularDepth/utils.py:160: patchs ==> patches, paths
./BoostingMonocularDepth/utils.py:171: patchs ==> patches, paths
./BoostingMonocularDepth/utils.py:184: patchs ==> patches, paths
./BoostingMonocularDepth/utils.py:185: patchs ==> patches, paths
./BoostingMonocularDepth/utils.py:186: patchs ==> patches, paths
./BoostingMonocularDepth/README.md:120: dependancies ==> dependencies
./BoostingMonocularDepth/README.md:181: evalution ==> evaluation, evolution
./BoostingMonocularDepth/README.md:182: accelarate ==> accelerate
./BoostingMonocularDepth/demo.py:24: wieghts ==> weights
./BoostingMonocularDepth/demo.py:213: patchs ==> patches, paths
./BoostingMonocularDepth/pix2pix/data/base_dataset.py:58: ususally ==> usually
./BoostingMonocularDepth/pix2pix/models/pix2pix4depth_model.py:12: orignal ==> original
./BoostingMonocularDepth/pix2pix/models/pix2pix4depth_model.py:155: udpate ==> update
./BoostingMonocularDepth/pix2pix/options/base_options.py:17: initailized ==> initialized
./BoostingMonocularDepth/pix2pix/util/visualizer.py:60: saveing ==> saving

The resulting depth images have more details but also worse depth estimation comparing with other MiDaS implementations, false color and even exr output match poorly with scene.

you have mentioned the code works properly with pytorch <=1.8 as as well, but torch.utils.ffi has been since 0.4.0 depreciated. In the requirements.txt, one sees pytorch==1.2

with torch 0.4.1, the following error occurs:
AttributeError: module 'torch' has no attribute 'ops'
or
File "./structuredrl/models/syncbn/modules/functional/_syncbn/_ext/syncbn/init.py", line 3, in
from ._syncbn import lib as _lib, ffi as _ffi
ImportError: No module named 'modules.functional._syncbn._ext.syncbn._syncbn'

I am running the LeRes and with input (3024, 5376, 3)
It tried to downscale by factor of 0.23809523809523808 and result in depthmap output of 720 * 1280
Is this any option that i can adjust to match the original size?

This project produces great results! However, I'm seeing some banding on the models when viewing them in 3d:

Original high res:
https://drive.google.com/file/d/1sDHva_euGTQnwx7Z54eEH9TuqKCGB5gq/view?usp=sharing

Merged:
https://drive.google.com/file/d/1DXgd_tcS13O6yHaGESFks2UlCh25PMYq/view?usp=sharing

Is the model trained on 8 bit inputs? It can output in 16 bit, but is there a way to avoid 8 bit processing in the middle, which could be reducing the quality of the final model?

(Edit: I've only looked at this through BoostYourOwnDepth)

!python run.py --Final --data_dir ./inputs --output_dir ./output1 --depthNet 0

Traceback (most recent call last):
File "run.py", line 498, in
run(dataset_, option_)
File "run.py", line 242, in run
tobemergedto[h1:h2, w1:w2] = np.multiply(tobemergedto[h1:h2, w1:w2], 1 - mask) + np.multiply(merged, mask)
ValueError: operands could not be broadcast together with shapes (672,671) (671,672)

so, add code in run.py (244 - 246)

            h, w = tobemergedto[h1:h2, w1:w2].shape
            mask = cv2.resize(mask, (w, h), interpolation=cv2.INTER_LINEAR)
            merged = cv2.resize(merged, (w, h), interpolation=cv2.INTER_CUBIC)

great work, would it be possible to make a demo for this on https://gradio.app/hub or would that be not allowed due to the license?

Link is not working ....
Can you please check it

Thanks for the great work! Is it still possible to use the boosted depth for shape recovery as in the original LeReS work? It looks like the output of base depth model goes through several normalization steps before merging.

I cannot understand this line in "evaluatedataset.m":

1- why "estimate_disp_ " should rescale between the "min_gt_disp" and 1.

2- And what is the disparity and depth's unit? are those in meters?

Any help would be appreciated.

Hallo, firstly thanks for the great work.
I have been running mainly Midas in Google Colab with great success, but today it won't run and throws up this error
FileNotFoundError: [Errno 2] No such file or directory: 'midas/model.pt'

Leres runs fine, but for some reason, I can't get it to run Midas, any ideas?
Thanks

Hi, I was wondering if the output is only a disparity map or it is possible to obtain a metric depth map using this network.
Thank you in advance!

Hello,

Thank you for building this model. Is the licensing fine with running BMD on an image dataset and distributing resulting depths?

Fantastic work! Thanks for sharing. Is there any possibility you might consider creating a google colab notebook for the less technically inclined like myself to try this out?

Thought to give my new GPU test on this mixed with 3DP but noticed some suspicious slowness, a process that should have taken 10 minutes was taking 7 hours. I worked down the list of why and thought to test this on its own, and sure enough, it provided the information I needed. It gives a warning that the RTX 3070 in my system has a newer compute version than is supported, and instead of defaulting to my RTX 2060 (this is supported and works in seconds), it defaults to a single CPU thread (1/16).

The same issue is found in 3DP and I'll have to forward the issue there too.

conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c conda-forge

on the pytorch page makes the card work and the results are not altered by the change

Is there is particular reason to use Midas 2 over Midas 3?

I was wondering if we can use this for video input or on real-time video from a webcam?

Hallo.
I have had this error pop up over the last couple of days while running Colab. I have Factory reset runtime, and terminated previous sessions.
Any help will be much appreciated, thanks in advance, Jonathan

Cloning into 'BoostingMonocularDepth'...
remote: Enumerating objects: 298, done.
remote: Counting objects: 100% (298/298), done.
remote: Compressing objects: 100% (231/231), done.
remote: Total 298 (delta 115), reused 217 (delta 58), pack-reused 0
Receiving objects: 100% (298/298), 55.49 MiB | 3.57 MiB/s, done.
Resolving deltas: 100% (115/115), done.
Access denied with the following error:

Cannot retrieve the public link of the file. You may need to change
the permission to 'Anyone with the link', or have had many accesses. 

You may still be able to access the file from the browser:

 https://drive.google.com/u/0/uc?id=1cU2y-kMbt0Sf00Ns4CN2oO9qPJ8BensP 

mv: cannot stat 'latest_net_G.pth': No such file or directory
Access denied with the following error:

Cannot retrieve the public link of the file. You may need to change
the permission to 'Anyone with the link', or have had many accesses. 

You may still be able to access the file from the browser:

 https://drive.google.com/uc?id=1nqW_Hwj86kslfsXR7EnXpEWdO2csz1cC 

mv: cannot stat 'model.pt': No such file or directory
--2022-02-18 10:13:33-- https://cloudstor.aarnet.edu.au/plus/s/lTIJF4vrvHCAI31/download
Resolving cloudstor.aarnet.edu.au (cloudstor.aarnet.edu.au)... 202.158.207.20
Connecting to cloudstor.aarnet.edu.au (cloudstor.aarnet.edu.au)|202.158.207.20|:443... connected.
HTTP request sent, awaiting response... 200 OK
Syntax error in Set-Cookie: 5230042dc1897=jdrh0u0lfe457ebu190haj24n7; path=/plus;; Secure at position 53.
Syntax error in Set-Cookie: oc_sessionPassphrase=tpTKJ%2B%2BsHMKirt%2BEZ4dF2UZOMzNclIR4jgr1AUEX%2BkxgZQbBKzvB9Z2qkVnETdoOSIE66u0s1vRyR%2BzLGfx7GdwhHYmpRShSJmtNYAKETV9YeL9wBN%2FQgPbYLIx364TS; path=/plus;; Secure at position 174.
Length: 530760553 (506M) [application/octet-stream]
Saving to: ‘download’

download 100%[===================>] 506.17M 3.85MB/s in 2m 3s

2022-02-18 10:15:38 (4.12 MB/s) - ‘download’ saved [530760553/530760553]

Hi
Could the local boosting process (creating, selecting and merging the patches) be done in parallel?

Hey,
Thanks for sharing this code, your work is incredible.

I successfully ran Midas and ResNet but when I try to run SRGNet, I have an issue with a file that seems not to exist:

I simply run
python run.py --Final --data_dir IN --output_dir OUT --depthNet 1

And I obtain

    from structuredrl.models import DepthNet
  File "/home/conthe/Schreibtisch/merge/BoostingMonocularDepth-main/structuredrl/models/DepthNet.py", line 18, in <module>
    from modules import nn as NN
  File "./structuredrl/models/syncbn/modules/nn/__init__.py", line 1, in <module>
    from .syncbn import *
  File "./structuredrl/models/syncbn/modules/nn/syncbn.py", line 21, in <module>
    from modules.functional import batchnorm2d_sync
  File "./structuredrl/models/syncbn/modules/functional/__init__.py", line 1, in <module>
    from .syncbn import batchnorm2d_sync
  File "./structuredrl/models/syncbn/modules/functional/syncbn.py", line 18, in <module>
    from ._syncbn._ext import syncbn as _lib_bn
  File "./structuredrl/models/syncbn/modules/functional/_syncbn/_ext/syncbn/__init__.py", line 3, in <module>
    from ._syncbn import lib as _lib, ffi as _ffi
ModuleNotFoundError: No module named 'modules.functional._syncbn._ext.syncbn._syncbn' ```

Once again, thousands thanks!!

Hi! Thank you a lot for your work, the results I get with it are really great!
I have a question however: is it possible to speed up the inference somehow?
I have a dataset of 40k images of resolution 256x256 which I need to extract the depth maps for.
I launched the inference with LeRes with the command you specified without changing the code:

python run.py --Final --data_dir <input_path> --output_dir <output_path> --depthNet 2

and after 9 hours on A100 80GB, it computed the depths maps for only 3K out of 40K images.
So, it will take me 5 days on A100 80GB to do inference for my dataset.
Is it possible to speed up this process somehow?

Thats really Awesome Work. The results look really great, Thanks a lot for sharing it.

I tried the code from Git Repo on some other images and the depth maps obtained look really awesome. I had one question though, How can I get actual depth values from the depth map obtained from the model?

Hi,
I faced a problem during checking both LeRes and MiDas output of this command:
python run.py --colorize_results --Final --data_dir PATH_TO_INPUT --output_dir PATH_TO_RESULT --depthNet #[0,1 or 2]

The color range of output images is different during running depthNet 0 and 2 separately. see below please:
I would appreciate it if you could guide me on why this happened? it seems the colors become opposite and it affects the evaluation result, too.

Hi, I'm reading your exciting paper and code, but I get a little confused about the mergenet part.

1. It seems that the depth ground truth when training the mergenet is generated by passing the image data in $root_dir/ibims1/rgb and $root_dir/middleburry/rgb to a pretrained MiDas, and use a matlab file(create_crops.m) to save the results in gt_fake_result_dir. But the inner data, which is used as the input of the mergenet, is also made by the same pretrained midas network with the same resolution, so what makes the gtfake more accurate than the inner?
2. It seems the mergenet has two different functions: merge the whole-image estimations with low and high resolutions and merge the base estimate(whole-image estimation) and the patch estimate(estimation of a part of the image). Is that right?

when i run the train pix2pix code, i get the error No module named 'pix2pix'

hi，I use it to test a image，but the color is different from the example，the near thing is black or purple，the right thing is yellow，can u tell me why？Thanks very much！！！

Can you share name of some tools that were used to create your explanation video. I found it very interesting.
Thankyou.

Hey,

I was following the instructions for downloading the datasets needed for re-training the merging network, but the link to the Ibims1-core-raw dataset does not seem to exist anymore.

Thanks.

The results look awesome. However the result of processing image sequence from video introduces noticeable temporal inconsistency.
Any plan to implement some solution to restore temporal consistency, or may be you can suggest any work around to solve that?

https://github.com/huggingface/accelerate <<< integration options for multi tpu approach ??

Hi all, I have successfully cloned the repository and run it on my Windows PC a few months ago. I am trying it on my Linux device now but I encounter an HTTP Error. I run using this:
python run.py --Final --data_dir PATH_TO_INPUT --output_dir PATH_TO_RESULT --depthNet 0
I created a conda environment with python=3.7 and installed the following:
conda install pytorch torchvision opencv cudatoolkit=10.2 -c pytorch
conda install matplotlib
conda install scipy
conda install scikit-image

with the --data_dir and --output_dir inserted and the exact error message I obtained is this:

initialize network with normal
loading the model from ./pix2pix/checkpoints/mergemodel/latest_net_G.pth
Loading weights: midas/model.pt
Downloading: "https://github.com/facebookresearch/WSL-Images/archive/master.zip" to /home/zhenkai/.cache/torch/hub/master.zip
Traceback (most recent call last):
File "run.py", line 580, in <module>
run(dataset_, option_)
File "run.py", line 59, in run
midasmodel = MidasNet(midas_model_path, non_negative=True)
File "/home/zhenkai/boosting_monocular/BoostingMonocularDepth/midas/models/midas_net.py", line 30, in __init__
self.pretrained, self.scratch = _make_encoder(features, use_pretrained)
File "/home/zhenkai/boosting_monocular/BoostingMonocularDepth/midas/models/blocks.py", line 6, in _make_encoder
pretrained = _make_pretrained_resnext101_wsl(use_pretrained)
File "/home/zhenkai/boosting_monocular/BoostingMonocularDepth/midas/models/blocks.py", line 26, in _make_pretrained_resnext101_wsl
resnet = torch.hub.load("facebookresearch/WSL-Images", "resnext101_32x8d_wsl")
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/site-packages/torch/hub.py", line 345, in load
repo_dir = _get_cache_or_reload(github, force_reload, verbose)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/site-packages/torch/hub.py", line 144, in _get_cache_or_reload
download_url_to_file(url, cached_file, progress=False)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/site-packages/torch/hub.py", line 379, in download_url_to_file
u = urlopen(req)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 222, in urlopen
return opener.open(url, data, timeout)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 531, in open
response = meth(req, response)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 641, in http_response
'http', request, response, code, msg, hdrs)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 563, in error
result = self._call_chain(*args)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 503, in _call_chain
result = func(*args)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 755, in http_error_302
return self.parent.open(new, timeout=req.timeout)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 531, in open
response = meth(req, response)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 641, in http_response
'http', request, response, code, msg, hdrs)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 569, in error
return self._call_chain(*args)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 503, in _call_chain
result = func(*args)
File "/home/zhenkai/anaconda3/envs/boosting/lib/python3.7/urllib/request.py", line 649, in http_error_default
raise HTTPError(req.full_url, code, msg, hdrs, fp)
urllib.error.HTTPError: HTTP Error 404: Not Found

Feels like the problem lies with Downloading:"https://github.com/facebookresearch/WSL-Images/archive/master.zip". Any help is greatly appreciated. thank you!

Update: Tried the google collab file as well, it gives a similar error.

Hi, thank you for exciting model!

I'm trying to understand the code to implement it in other language.
I have some questions about the code.

1. The Interpolation method used for resize operation throughout the code are INTER_CUBIC, is it better than other interpolation methods for your method?
2. When doing double estimate, monocular depth estimation output is firstly resized to original image size and then resized to pix2pix model input size. Should the monocular depth estimation output be resized to pix2pix model directly? Is there performance difference between them?
3. Why is the line necessary?
4. Why is the line use patch_whole_estimate_base instead of patch_whole_estimate_updated?

How to improve FPS? I'm in 2K image, 2fps