--- Semantic Segmentation of Multi/Single-object Images with Auto Data Annotation
Data Source
characteristics: 4000 multiobject images, 1000000 single object images. 204 unique categories
Prerequisite
install all required packages as defined in Auto Annotation, Yolo-v7 for Segmentation
install labelme
cd labelme-main
pip install .
Methods
Step 1: use roboflow, a online data annotation tool to manually annotation 200-250 multi-object images. To get better result, use build-in data augmentation tool.
Step 2: train Auto Annotation to fit the manually labeled data.
python3 customTrain.py train --dataset=YourDatasetDir
Step 3: Use Auto Annotation to annotate data.
python3 annotate.py annotateCustom --image_directory=DatasetDirMissingLabel --label=CustomLabel --weights=XXX.h5 --displayMaskedImages=False
Step 4: Clean the data. Use the modified labelme to delete annotation files with obvious errors or fix it.
Step 5: With the new data, repeat Step 2-4 until you get satisfactory segmentation results.
Step 5.5: Download Pretrained Model here.
Step 6: Run data format conversion script in haihua_data_format_conversion.ipynb, which also automatically assign a classification label to each object according to ground truth label provided in Haihua dataset.
Step 6.1: Optionally, you are encouraged to manually assign a classification label using the provided labelme. Labels for some objects are missing in Haihua dataset for unknown reason
Step 8: Train your yolo-v7-segmentation model
