The link of Stereo Depth Estimation for Cityscapes is unavailable, can you fix it?

when I use script llike

CUDA_VISIBLE_DEVICES=0 python3 -u trainUDA_gta.py --config ./configs/configUDA_gta2city.json --name UDA-gta --resume /saved/DeepLabv2-depth-gtamono-cityscapestereo/05-03_02-13-UDA-gta/checkpoint-iter95000.pth | tee ./gta-corda.log

It would run again but the new checkpoint would be saved.

Thanks a lot for your fantastic work. When I followed your depth estimation mentioned in issue#7, I went to the https://playing-for-benchmarks.org. However,its camera calibration doesn't include intrinsic matrix directly, which is needed in Monodepth2 depth estimation. Would you kindly share the intrinsic of GTA you used in depth estimation? Or may I know a way to convert GTA's projection matrix to intrinsic matrix?

Thanks for sharing the code, and it brings the amazing improvement for this filed.

I notice that you have used backbone with parameters pretrained on MSCOCO which is the same with DACS, and have you tried backbone pretrained on ImageNet? If yes, could you please provide the corresponding results?

When I revisited the performance of your GTA2City, I found that the MIOU could only reach about 54.8 after 250,000 iterations. I didn't change anything except the 10.2 version of CUDA. Could you please provide the training log of your GTA2City?
Thanks a lot！！

I follow you code, and train a model. But it results may not meet the need.

I eval the model you share .

bash shells/eval_syn2city.sh

your share model.
in syn2city : 19 classes : meanIout: 0.4771
I train the model:
in syn2city : 19 classes : meanIout: only: 0.46.7

in the train. I find the warning, So I want to know if it may cause the result drop.

/home/ailab/anaconda3/envs/yy_CORDA/lib/python3.7/site-packages/torch/optim/sgd.py:68: UserWarning: optimizer contains a parameter group with duplicate parameters; in future, this will cause an error; see github.com/pytorch/pytorch/issues/40967 for more information
  super(SGD, self).__init__(params, defaults)
D_init tensor(134.8489, device='cuda:0', grad_fn=<DivBackward0>) D tensor(134.5171, device='cuda:0', grad_fn=<DivBackward0>)

I can't download the checkpoints file from your links, when click into the google drive, The file size is shown to be 2GB, but it was only 0B when downloaded

In your paper use DACS as baseline.
However, DACS don't propose the SYNTHIA->CITYSCAPES trainning.
So, I want to ask you how you use trainning DACS.
I read you code. Should I need to delete all code which use depth in trainning. because I need to use DACS in SYNTHIA->CITYSCAPES to obtain data in my work

Another question, When trainning finished, predictions direction will be created and some test pictures saved here.Where Can I find the code about it.

Maybe it's too long because this work is finished two years ago.You may forget some details.
I will be very appreciated if you reply me !

if the forward code, return out ,an extra space symbol

   def forward(self, x):
        out = self.conv2d_list[0](x)
        for i in range(len(self.conv2d_list)-1):
            out += self.conv2d_list[i+1](x)
            return out

this is the code in your code

class Classifier_Module(nn.Module):

    def __init__(self, dilation_series, padding_series, num_classes):
        super(Classifier_Module, self).__init__()
        self.conv2d_list = nn.ModuleList()
        for dilation, padding in zip(dilation_series, padding_series):
            self.conv2d_list.append(nn.Conv2d(256, num_classes, kernel_size=3, stride=1, padding=padding, dilation=dilation, bias = True))

        for m in self.conv2d_list:
            m.weight.data.normal_(0, 0.01)

    def forward(self, x):
        out = self.conv2d_list[0](x)
        for i in range(len(self.conv2d_list)-1):
            out += self.conv2d_list[i+1](x)
            return out

I this this forward is possible.beaceuse your code, use list contain four elements,if return out have space, this may do only twice without fourth

   def forward(self, x):
        out = self.conv2d_list[0](x)
        for i in range(len(self.conv2d_list)-1):
            out += self.conv2d_list[i+1](x)
        return out

   self.__make_pred_layer(Classifier_Module,[6,12,18,24],[6, 12,18, 24],NUM_OUTPUT[task]

   def _make_pred_layer(self,block, dilation_series, padding_series,num_classes):
        return block(dilation_series,padding_series,num_classes)

I download your pretrained model, and start demo
But I find train iou 0.0

(yy_corda) ailab@ailab:/media/ailab/data/yy/corda$ bash shells/eval_gta2city.sh
./checkpoint/gta
Found 500 val images
Evaluating, found 500 batches.
100 processed
200 processed
300 processed
400 processed
500 processed
class  0 road         IU 94.81
class  1 sidewalk     IU 62.18
class  2 building     IU 88.03
class  3 wall         IU 33.09
class  4 fence        IU 43.51
class  5 pole         IU 39.93
class  6 traffic_light IU 49.46
class  7 traffic_sign IU 54.68
class  8 vegetation   IU 88.01
class  9 terrain      IU 47.67
class 10 sky          IU 89.22
class 11 person       IU 68.22
class 12 rider        IU 39.21
class 13 car          IU 90.25
class 14 truck        IU 51.43
class 15 bus          IU 58.37
class 16 train        IU 0.00
class 17 motorcycle   IU 40.38
class 18 bicycle      IU 57.42
meanIOU: 0.5767768805758403

I train my model on it, and test eval_syn2city.py. Here are 3 classes Iou 0.0 because missing classed in source domain.
but I download pretrained model ,and run eval_gta2city.sh
still miss one class train.
So, I want to know why. Is it may train class didn't appear city datasets？ So it's IOU is 0.

From what I understand after reading your paper, you do not need ground truth label data on the target domain to train the pseudo labels. However, when I look at cityscapes_loader, it seems I need to supply the ground truth seg maps as well.

I am trying to train the network on a custom dataset (that only depth maps, and ground truth seg map only on the source domain), but it looks I cannot get away without providing it. Do you have any thoughts on this?

Hi, thanks for your great work!

It would be great if you can elaborate more on how you obtain the monocular depth estimation.

I understand that you've uploaded the dataset, but it would be really helpful if I know exactly how you've done it.

From your paper, in the ablation study part: "We would like to highlight that for both stereo and monocular depth estimations, only stereo pairs or image sequences from the same dataset are used to train and generate the pseudo depth estimation model. As no data from external datasets is used, and stereo pairs and image sequences are relatively easy to obtain, our proposal of using self-supervised depth have the potential to be effectively realized in real-world applications."

So I image you get your monocular depth pseudo ground truth by:

1. Downloading target domain videos (here Cityscapes. Btw, where do you get Cityscapes videos?)
2. Train a Monodepth2 model on those videos (for how long?)
3. Use the model to get pseudo ground truth
   Then repeat to the source domain (GTA 5 or Synthia)

Am I getting it right? And is there any more important points you want to highlight when calculating such depth labels?

Regards,
Tu

Hello, nice work but i meet some question.

in 'data/cityscapes_loader.py' line 181-183:

depth = cv2.imread(depth_path, flags=cv2.IMREAD_ANYDEPTH).astype(np.float32) / 256. + 1.
if depth.shape != lbl.shape:
depth = cv2.resize(depth, lbl.shape[::-1], interpolation=cv2.INTER_NEAREST)
Monocular depth: in disparity form 0 - 65535

(1) Why the depth is calculated from x/256+1
(2) is it the depth or the disparity ? In the official doc of cityscapes, it say disparity = (x-1)/256

Thank you!