zhkkke / pixelssl Goto Github PK

Hi there, thanks for your work, it's very inspiring!

And now I want to use the job in my project, but in 3D. I found that the FlawDetector for 2D is stacked of some conv layers with kernel size is 4 stride is 1 or 2 or some stuff.

But my input size is 256, 256 after the self.conv3_1 will cause errors. So I have to modify kernel size from 4 to 3, and now before interpolating the feature map, the x's shape is (1, 1, 8, 8, 8), but to interpolating to shape of (1, 1, 16, 256, 256), the gap between the x and the task_pred seems too large.

But in 2D mode, I set the input is (3, 256, 256) while the num_classes is 14, the x will be interpolated from (1, 1, 8, 8) to (1, 1, 256, 256). Is is reasonable?

Thanks a lot!

Hi ZHKKKE,

First of all, thank you for your work. Currently, I retrain the gct by PSPNet with the ResNet-101 backbone in Pascal VOC, and use the parameter of im_size=513, batch_size=4 with 4 gpus. However, i am getting the error of insufficient memory. I retrained other methods you offered by using the parameter of im_size=513, batch_size=4 with 4 gpus and can get the accuracy provided by README.md.

I want to know how you train the gct with 4 GPUs? Save memory by changing im_size=513 to im_size=321？Or is there any other way？

Thank you and regards

great work. Thanks for your amazing codebase.
I have some questions about this paper "Guided Collaborative Training for Pixel-wise
Semi-Supervised Learning"

1.I'm wondering if I can just use max score of a pixel as an evaluation criterion without Flaw Detector in semantic segmentation task? If so, how would it work if I use score directly, have you ever done such experiment?

2. Is Flaw Correction Constraint forcing the error to 0 to correct the result of semantic segmentation? This loss, not quite understand what it means.

Dear author:
I have a question about the inference setting. In this section:

    def _val_prehandle(self, image, label):
        sample = {self.IMAGE: image, self.LABEL: label}
        composed_transforms = transforms.Compose([
            FixScaleCrop(crop_size=self.args.im_size),
            Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
            ToTensor()])

        transformed_sample = composed_transforms(sample)

        return transformed_sample[self.IMAGE], transformed_sample[self.LABEL]

I find that you crop the image as the input and calculate the metrics on the cropped image. However, I think we should use the whole image to calculate the metric. Based on this setting, the supervised full baseline is 2~3% mIoU lower than the raw performance. Could you explain it?

Hi, ZHKKKe, Thank you for your excellent code.

I found a suspected bug in task/sseg/func.py.

In the function metrics, you reset all meters named acc_str/acc_class_str/mIoU_str/fwIoU_str.
if meters.has_key(acc_str): meters.reset(acc_str) if meters.has_key(acc_class_str): meters.reset(acc_class_str) if meters.has_key(mIoU_str): meters.reset(mIoU_str) if meters.has_key(fwIoU_str): meters.reset(fwIoU_str)
When I test your pre-trained model deeplabv2_pascalvoc_1-8_suponly.ckpt, I found the Validation metrics logging the whole confusion matrix. Shouldn‘t we count the single image acc/mIoU independently?

I'm not sure whether my speculation is right, could you help me?

Dear author:
Thank you for sharing! Could you share more data splits of your ECCV paper, such as data split of 1/16, 1/4, 1/2 of VOC? We want to run experiments based on more splits and make a comparison with the numbers reported in the paper. Thank you!

Hello, thanks for your perfect work. I have a question about the performance of PSPNet , when i use PSPNet alone in my own dataset and my own code and trainning with 1/2 samples, the miou could reach about 68%. But when I change to your code and trainningwith suponly, the miou is only 60% . Could you please tell me what may be the reason for this.

Thanks for your sharing and the repo is quite helpful for me to understand the work in SSL segmentation. If possible, could you add the implementation of Semi-supervised Semantic Segmentation via Strong-weak Dual-branch Network (ECCV 2020), which is a simply dual branch network. It's a quite easy and inituitive idea but I could not reproduce the results with deeplabv2. It would be great if you could add this into the repo.

Hi, thanks for your nice work!

I have noticed that you only give us the data split of VOC2012, will you offer us the splits of cityscapes dataset?

And from your scripts,
The labeled data used in your experiments only samples in the order of names from the txt file,

Thanks for your great work for tackling the pixel-wise semi-supervised tasks. I am currently following it and I have the following question.

Should the returned value of 'out' at https://github.com/ZHKKKe/PixelSSL/blob/master/task/sseg/module/deeplab_v2.py#L85 be out of the for loop? Otherwise, the ASPP only adds the outputs of dilation rates 6 and 12.

Thanks in advance : )

In the GCT segmentation task, I see that the validation metrics include the left and right models, more specifically l-metric-mIOU and r-metric-mIOU. Which model's metric should I use to compare with other baselines such as self-training, self-training++?