Closed LPXTT closed 2 years ago
Besides, I am new to semantic segmentation task and don't know how to preprocess cityscape data. Could you please give me some instructions?
Thanks for your interest.
The download link for the pre-trained model is same as the CPS, and you can find them in here. Please don't worry about those bucket_namespaces, they are for k8s & dgx-clustering launch. You can train our approach in the normal way.
The processed cityscapes can be found in this repository, and please note that the directory for the cityscapes follow this structure:
city_scape/
├── annotation
│ └── city_gt_fine
│ ├── train
│ └── val
└── images
└── city_gt_fine
├── train
└── val
Best, Yuyuan
Thank you!
发自我的iPhone
在 2022年8月4日,下午1:28,Yuyuan Liu @.***> 写道:
Thanks for your interest.
The download link for the pre-trained model is same as the CPS, and you can find them in here. Please don't worry about those bucket_namespaces, they are for k8s & dgx-clustering launch. You can train our approach in the normal way.
The processed cityscapes can be found in this repository, and please note that the directory for the cityscapes follow this structure:
city_scape/ ├── annotation │ └── city_gt_fine │ ├── train │ └── val └── images └── city_gt_fine ├── train └── val Best, Yuyuan
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.
Sorry. I didn't find the processed cityscapes in this repository. There are only original files not processed files.
Hi Yuyuan, How can I get the SegmentationClassAug? I found SegmentationClass and SegmentationClassAug in vocsplit/*.txt. While, there is no SegmentationClassAug in pascalVOC.
@LPXTT For the Cityscapes, the process stage only transfers the label classes. It contains around 30 classes in the original setting, but due to the long-tail issue, we only utilise 19 of them for training. The rest categories are treated as ignore index (255) so that what you are looking for is processed pixel-wise labels (that only contain 19 classes). The images should be as same as you download from the official website.
On the other hand, the download link for the SegmentationClassAug can be found in this repository. For a brief explanation, the Pascal VOC 12 contains around ~1400 official (high-quality) labels and ~9000 unofficial low-quality labels. The SegmentationClassAug should contain those low-quality labels.
Hope that helps!
Best, Yuyuan
Thank you!
Hello author, could you provide the download link of ResNet50 and ResNet101? I don't know the bucket_namespace and bucket_name in the download script. I downloaded two pretrained models from other place but the size dosen't match. Thanks.