Aryan-Mishra24
commented
1 year ago To successfully convert the Cityscapes dataset while filtering specific classes such as traffic lights and poles, you can follow the steps below:
First, clone the Cityscapes scripts repository: bash
git clone https://github.com/mcordts/cityscapesScripts.git cd cityscapesScripts/cityscapesscripts/preparation Modify the createTrainIdInstanceImgs.py script to filter out specific classes. In the createTrainIdInstanceImgs.py file, find the following lines: python
Load the annotation
objects = csHelpers.loadAnnotaton(labelImgPath) Add a filtering function after the lines mentioned above: python
def filter_objects(objects, classes_to_keep): filtered_objects = [] for obj in objects: if obj.label.name in classes_to_keep: filtered_objects.append(obj) return filtered_objects
classes_to_keep = ['traffic light', 'pole'] objects = filter_objects(objects, classes_to_keep) Replace the classes_to_keep list with the classes you want to keep in the final annotations.
In the json2instanceImg.py script, find the following line: python
if not anno['regions']: Replace the above line with the following: python
if len(anno['regions']) == 0: Now you can run the createTrainIdInstanceImgs.py script to generate the filtered dataset: bash
python createTrainIdInstanceImgs.py --cityscapesPath path/to/cityscapes --outputPath path/to/output Finally, run the json2instanceImg.py script to convert the modified annotations to JSON files: bash python json2instanceImg.py --cityscapesPath path/to/cityscapes --outputPath path/to/output After following these steps, the output JSON files should contain only the filtered classes (e.g., traffic lights and poles). Make sure to replace path/to/cityscapes with the path to your Cityscapes dataset and path/to/output with your desired output path.
Has somebody successfully converted cityscapes dataset filtering classes such as traffic lights, poles, etc? Indeed, I did the change of desired classes in _categoryinstancesonly and also commented on the invalid contour warning. However, even if I did that, I still get .json output files with 0 annotations after loading all the images.