Inference on single stereo image pairs

[kHarshit]

Thanks for your work.

I want to perform inference on a single stereo image pair. This repo only contains evaluation on KITTI dataset images.

[CMartinETH]

I am working on the exact same thing. Right now, I am trying to write a function that stores the inference results in an image. But my results are always blurry (see image below).

Within the module finetune.py, I use the test function and use the output and convert it to a PIL image.
Maybe someone has an explanation for this.

[JF-Lee]

I am working on the exact same thing. Right now, I am trying to write a function that stores the inference results in an image. But my results are always blurry (see image below).
Within the module finetune.py, I use the test function and use the output and convert it to a PIL image.
Maybe someone has an explanation for this.

Hi，I also encountered this problem，and I use torchvision.utils.save_image to save the tensor, but get blank pic.

[CMartinETH]

@JF-Lee I found a solution to the issue: I wrote a function that stores the output of the network (keep in mind, you get 4 different images, one image per stage multiplied by the batch size - so say you have a batch size of 8, you get 8 images in one tensor as output, 4 times).
This can be done from within the test function. I hand over the output and x (to know which is the current stage I am saving the image of, or to only save the image of stage 3).

Then the code in the separate file looks something like this:

for i in range(output.size()[0]):
        img_cpu = np.asarray(output.cpu())
        img_save = np.clip(img_cpu[i, :, :], 0, 2**16)
        img_save = (img_save * 256.0).astype(np.uint16)
        name = "some path name"
        cv2.imwrite(name, img_save)

I hope that helps!

[JF-Lee]

Oh, It's great, thank you for your reply!!!

…

------------------&nbsp;原始邮件&nbsp;------------------ 发件人:&nbsp;"CMartinETH"<notifications@github.com&gt;; 发送时间:&nbsp;2020年6月3日(星期三) 晚上9:18 收件人:&nbsp;"mileyan/AnyNet"<AnyNet@noreply.github.com&gt;; 抄送:&nbsp;"走走停停ゝ"<445099619@qq.com&gt;;"Mention"<mention@noreply.github.com&gt;; 主题:&nbsp;Re: [mileyan/AnyNet] Inference on single stereo image pairs (#12) @JF-Lee I found a solution to the issue: I wrote a function that stores the output of the network (keep in mind, you get 4 different images, one image per stage multiplied by the batch size - so say you have a batch size of 8, you get 8 images in one tensor as output, 4 times). This can be done from within the test function. I hand over the output and x (to know which is the current stage I am saving the image of, or to only save the image of stage 3). Then the code in the separate file looks something like this: for i in range(output.size()[0]): img_cpu = np.asarray(output.cpu()) img_save = np.clip(img_cpu[i, :, :], 0, 2**16) img_save = (img_save * 256.0).astype(np.uint16) name = "some path name" cv2.imwrite(name, img_save) I hope that helps! — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[mileyan]

@CMartinETH, thanks too much for helping solve this issue. I will update this function to the master branch.

[lawo123]

Why is my prediction so bad？

[namoamituo]

@JF-Lee I found a solution to the issue: I wrote a function that stores the output of the network (keep in mind, you get 4 different images, one image per stage multiplied by the batch size - so say you have a batch size of 8, you get 8 images in one tensor as output, 4 times). This can be done from within the test function. I hand over the output and x (to know which is the current stage I am saving the image of, or to only save the image of stage 3).

Then the code in the separate file looks something like this:

for i in range(output.size()[0]):
        img_cpu = np.asarray(output.cpu())
        img_save = np.clip(img_cpu[i, :, :], 0, 2**16)
        img_save = (img_save * 256.0).astype(np.uint16)
        name = "some path name"
        cv2.imwrite(name, img_save)

I hope that helps!

hi，Could you show me some more details? I'm not sure where the above code should be placed in the test function

[leizhenyu-lzy]

Why is my prediction so bad？

got similar result, it looks a little blurred

[leizhenyu-lzy]

@JF-Lee I found a solution to the issue: I wrote a function that stores the output of the network (keep in mind, you get 4 different images, one image per stage multiplied by the batch size - so say you have a batch size of 8, you get 8 images in one tensor as output, 4 times). This can be done from within the test function. I hand over the output and x (to know which is the current stage I am saving the image of, or to only save the image of stage 3).
Then the code in the separate file looks something like this:

for i in range(output.size()[0]):
        img_cpu = np.asarray(output.cpu())
        img_save = np.clip(img_cpu[i, :, :], 0, 2**16)
        img_save = (img_save * 256.0).astype(np.uint16)
        name = "some path name"
        cv2.imwrite(name, img_save)

I hope that helps!

hi，Could you show me some more details? I'm not sure where the above code should be placed in the test function

you can put it in the test() function in the file finetune.py

def test(dataloader, model, log):

    stages = 3 + args.with_spn
    D1s = [AverageMeter() for _ in range(stages)]
    length_loader = len(dataloader)

    model.eval()

    for batch_idx, (imgL, imgR, disp_L) in enumerate(dataloader):
        imgL = imgL.float().cuda()
        imgR = imgR.float().cuda()
        disp_L = disp_L.float().cuda()

        with torch.no_grad():
            outputs = model(imgL, imgR)
            print(f"[test] : len(outputs)={len(outputs)}")  # 4
            for x in range(stages):
                output = torch.squeeze(outputs[x], 1)
                print(f"[test] : output.shape={output.shape}")  # torch.Size([8, 1, 368, 1232])
                for i in range(output.size()[0]):  # 8
                    img_cpu = np.asarray(output.cpu())
                    img_save = np.clip(img_cpu[i, :, :], 0, 2**16)
                    img_save = (img_save * 256.0).astype(np.uint16)
                    name = "/home/lzy/Project/DepthEstimateBasedOnDeepLearning/AnyNet/results/test/"+str(x)+"_"+str(i)+".png"
                    cv2.imwrite(name, img_save)
                D1s[x].update(error_estimating(output, disp_L).item())

[Xi-Gong]

Why is my prediction so bad？

Compare to the paper result, looks like this is actually the best result AnyNet can get.