albert100121 / 360sd-net

Pytorch implementation of ICRA 2020 paper "360° Stereo Depth Estimation with Learnable Cost Volume"

360SD-Net

project page | paper | dataset

This is the implementation of our ICRA 2020 paper "360° Stereo Depth Estimation with Learnable Cost Volume" by Ning-Hsu Wang

Overview

How to Use

• Setup a directory for all experiments. All you have to do in advance may look like this,

# SETUP REPO
>> git clone https://github.com/albert100121/360SD-Net.git
>> cd 360SD-Net
>> mkdir output
>> cd conda_env
>> conda create --name 360SD-Net python=2.7
>> conda activate 360SD-Net
>> conda install --file requirement.txt

# DOWNLOAD MP3D Dataset
>> cd ./data
# reqest download MP3D Dataset
>> unzip MP3D Dataset
# request download SF3D Dataset
>> unzip SF3D Dataset

• Setup data and directories (opt to you as long as the data is linked correctly). Set the directory structure for data as follows:

# MP3D Dataset
./data/
     |--MP3D/
                 |--train/
                       |--image_up/
                       |--image_down/
                       |--disp_up/
                 |--val/
                       |--image_up/
                       |--image_down/
                       |--disp_up/
                 |--test/
                       |--image_up/
                       |--image_down/
                       |--disp_up/
# SF3D Dataset
./data/
     |--SF3D/
                 |--train/
                       |--image_up/
                       |--image_down/
                       |--disp_up/
                 |--val/
                       |--image_up/
                       |--image_down/
                       |--disp_up/
                 |--test/
                       |--image_up/
                       |--image_down/
                       |--disp_up/

• Training procedure:

# For MP3D Dataset
>> python main.py --datapath data/MP3D/train/ --datapath_val data/MP3D/val/ --batch 8

# For SF3D Dataset
>> python main.py --datapath data/SF3D/train/ --datapath_val data/SF3D/val/ --batch 8 --SF3D

• Testing prodedure:

# For MP3D Dataset
>> python testing.py --datapath data/MP3D/test/ --checkpoint checkpoints/MP3D_checkpoint/checkpoint.tar --outfile output/MP3D

# For SF3D Dataset
>> python testing.py --datapath data/SF3D/test/ --checkpoint checkpoints/SF3D_checkpoint/checkpoint.tar --outfile output/SF3D

# For Real World Data
>> python testing.py --datapath data/realworld/ --checkpoint checkpoints/Realworld_checkpoint/checkpoint.tar --real --outfile output/realworld

# For small inference
>> python testing.py --datapath data/inference/MP3D/ --checkpoint checkpoints/MP3D_checkpoint/checkpoint.tar --outfile output/small_inference

• Disparity to Depth:

>> python utils/disp2de.py --path PATH_TO_DISPARITY

Notes

• The training process will cost a lot of GPU memory. Please make sure you have a GPU with 32G or larger memory.
• For testing, 1080Ti (12G) is enough for a 512 x 1024 image.

Synthetic Results

• Depth / Error Map

* Projected PCL

Real-World Results

• Camera Setting

* Real World Results

Citation

@article{wang2019360sdnet,
	title={360SD-Net: 360° Stereo Depth Estimation with Learnable Cost Volume},
	author={Ning-Hsu Wang and Bolivar Solarte and Yi-Hsuan Tsai and Wei-Chen Chiu and Min Sun},
	journal={arXiv preprint arXiv:1911.04460},
	year={2019}
}

360sd-net's Issues

Error

tensor.sub_(mean[:, None, None]).div_(std[:, None, None])
RuntimeError: The size of tensor a (4) must match the size of tensor b (3) at non-singleton dimension 0

Could you please help me?

About the LCV module

Hi, nice work and thank you for your sharing !
From my perspective, the pixel in equirectangular projection is proportional to the degree_disparity.
For a 2*pi×pi equirectangular projection map, degree_disparity=i/pi where i in pixel coordinates disparity.
Please correct me if I'm wrong.

check the test output image

I test the small inference,the output are "0.npy,1.npy,2.npy". how can i view the disparity image?

Can I run this on Thetza Z1 Images I have created?

I have been creating 360 panos for some time now with my drone and the theta z1 camera.. can I inject my images to get 3d depth maps out of it?

The data contained in 360SD-Net/data are not suit.

Hello, I download your sample images from MP3D and SF3D provided in this repository. But I found that the up-view and down-view pictures are almost the same in your folders. So there are no differences between them.
The same thing occur in realworld folder, too. Please check out in your convenience, and let me know if I am wrong.

Best.

PreTrained Model

Can you please provide the pretrained model as I do not have the resources to train this data set?

Learnable Cost Volume model

Can you help me with this?

Epoch: 0%| | 0/500 [00:17<?, ?it/s]
Traceback (most recent call last):
File "main.py", line 348, in
main()
File "main.py", line 286, in main
loss = train(imgU_crop, imgD_crop, disp_crop)
File "main.py", line 196, in train
output1, output2, output3 = model(imgU, imgD)
File "/ssoft/spack/arvine/v1/opt/spack/linux-rhel7-skylake_avx512/gcc-8.4.0/py-torch-1.6.0-43xbre3fdhzp6upz6mfe3jk6rpwt5uky/lib/python3.7/site-packages/torch/nn/modules/module.py", line 722, in _call_impl
result = self.forward(*input, **kwargs)
File "/ssoft/spack/arvine/v1/opt/spack/linux-rhel7-skylake_avx512/gcc-8.4.0/py-torch-1.6.0-43xbre3fdhzp6upz6mfe3jk6rpwt5uky/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py", line 155, in forward
outputs = self.parallel_apply(replicas, inputs, kwargs)
File "/ssoft/spack/arvine/v1/opt/spack/linux-rhel7-skylake_avx512/gcc-8.4.0/py-torch-1.6.0-43xbre3fdhzp6upz6mfe3jk6rpwt5uky/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py", line 165, in parallel_apply
return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])
File "/ssoft/spack/arvine/v1/opt/spack/linux-rhel7-skylake_avx512/gcc-8.4.0/py-torch-1.6.0-43xbre3fdhzp6upz6mfe3jk6rpwt5uky/lib/python3.7/site-packages/torch/nn/parallel/parallel_apply.py", line 85, in parallel_apply
output.reraise()
File "/ssoft/spack/arvine/v1/opt/spack/linux-rhel7-skylake_avx512/gcc-8.4.0/py-torch-1.6.0-43xbre3fdhzp6upz6mfe3jk6rpwt5uky/lib/python3.7/site-packages/torch/_utils.py", line 395, in reraise
raise self.exc_type(msg)
TypeError: Caught TypeError in replica 0 on device 0.
Original Traceback (most recent call last):
File "/ssoft/spack/arvine/v1/opt/spack/linux-rhel7-skylake_avx512/gcc-8.4.0/py-torch-1.6.0-43xbre3fdhzp6upz6mfe3jk6rpwt5uky/lib/python3.7/site-packages/torch/nn/parallel/parallel_apply.py", line 60, in _worker
output = module(*input, **kwargs)
File "/ssoft/spack/arvine/v1/opt/spack/linux-rhel7-skylake_avx512/gcc-8.4.0/py-torch-1.6.0-43xbre3fdhzp6upz6mfe3jk6rpwt5uky/lib/python3.7/site-packages/torch/nn/modules/module.py", line 722, in call_impl
result = self.forward(*input, **kwargs)
File "/work/vita/danial/360SD-Net/models/LCV_ours_sub3.py", line 187, in forward
refimg_fea.size()[3]).zero()).cuda()
TypeError: new(): argument 'size' must be tuple of ints, but found element of type float at pos 3

Is CUDA11.2 supported?

I'm trying to run the test script with your model and this command:
python testing.py --datapath data/realworld/ --checkpoint checkpoints/Realworld_checkpoint/checkpoint.tar --real --outfile output/realworld

but I faced this problem:

I wonder is cuda11.2 supported?

3d conv?

Some questions about reproduction.

Hi,

I am trying to run this code with our own synthetic dataset.

I rendered two equirectangular images 512x1024 with a vertical stereo setup as yours.

However, I got a questionalble result.

Belows are input_up_image, input_down_image, result_with_MP3D_ckpt, result_with_SF3D_ckpt, and result_with_Real_ckpt.
For the last one, I gave args.real = True.

Do you think I am running incorrectly? or are these expected results? Please let me know.

For more information, I ran testing.py, and I changed __imagenet_status in preprocess.py.

I didn't work with the original code:
__imagenet_stats = { 'mean': [0.485, 0.456, 0.406], 'std': [0.229, 0.224, 0.225] }

, so I added some numbers like
__imagenet_stats = { 'mean': [0.485, 0.456, 0.406, 1], 'std': [0.229, 0.224, 0.225, 1] }

and it didn't matter much which number I gave. Should I change these numbers to reproduce your performance?

Problems encountered while downloading the dataset

Hi,
I received this email, and then downloaded the data set according to the first link,

but it looked like this after downloading,

May I ask whether I made a mistake? How to download the dataset correctly?

Thank you!