Closed Tranbaber closed 5 months ago
which file was used for the visualization above?
I used configs/Mobile_Seed/MS_tiny_cityscapes.py @martin-liao
I will check the visualization code later. The demo/image_demo.py
file is a good practice for inference and visualization. You might try it.
I have used demo/image_demo.py for inference and the config file is also configs/Mobile_Seed/MS_tiny_cityscapes.py and the inference comes out without any problem, but the above mentioned problem occurs when I use the .pth model file that I got from my own training for inference. So I think it might be a training issue. @martin-liao
Have you checked the semantic boundary label?
The semantic boundary label I use for my training looks like this:
The semantic boundary label generated according to the latest data_preprocess/cityscapes-preprocess/code/demoPreproc_gen_png_label.m looks like this: @martin-liao
it seems that everything is right…… could you achieve the miou performance reported in the paper?
Your model works very well, but I didn't reach the MIoU in the paper, probably due to equipment and dataset, etc. Here are the results of my training data:
Visualization results: @martin-liao
The mIoU is 69.96, significantly lower than 78.44 reported in the paper. Are you sure you haven't made any modifications to the configurations, such as reducing the batch size?
I have noticed memory=8052
in the provided screenshot. It means that you either crop the input image to a smaller size or reduce the batch size.
For the boundary map visualization, I will check it later.
Yes, sorry, I set samples_per_gpu=2 before training considering the GPU video memory, which is really an important reason that can affect the model performance. @martin-liao
I have checked the visualization code, and everything is right. binary boundary semantic boundary semantic
The semantic boundary label I use for my training looks like this:
The semantic boundary label generated according to the latest data_preprocess/cityscapes-preprocess/code/demoPreproc_gen_png_label.m looks like this: @martin-liao
background should be black (255) instead of white (0)
You really don't have a problem with your results. When I use demo/image_demo.py for inference, there is also no problem with visualizing the results if I use the weights file you trained on; however, if I use the weights file I trained on, nothing shows up in the bibound plot, it's blank. I think it's a problem with my training. Also, I would like to ask, the background of the semantic edge graph in the dataset you used for training is black right? Does it have to be black? @martin-liao
yeah,it's crucial
okay, thanks very much! @martin-liao
Since the problem has been resolved, I will close the issue.
Hello Author! I was able to get very good results in semantic segmentation and edge detection tasks after training, but during prediction, the bibound graph shows up blank and there is no effect graph like in the demo. What is the reason for this? Thank you for your answer.
bibound graph: