Based on SMA and YOLO V7 and Blender Simulation
The Earth and Satellite model used in this project are downloaded from Free3D.com
Based on the motion pattern of the satellite, I animated the model and output 1200 images, covering all the different attitudes.
ISAT_with_segment_anything is an awesome tool which use segment anything to realize semantic annotation.
But it's only output COCO Dataset format annotation, so I have written a conversion program to convert their annotation files to VOC format.
The conversion program is here.
After annotating can we finally training our own YOLO v7.
- Preparing Dataset This project using VOC format dataset for training.
Organizing your own Dataset like following structure:
- VOCdevkit
- VOC2007
- Annotation
- Your Annotation xmls
- JPEGImages
- Your Pictures
- Annotation
- VOC2007
- Progressing Dataset
run voc_annotation.py to obtain 2007_train.txt and 2007_val.txt for training。
change classes_path in voc_annotation.py.
classes_path refer to ./cls_classes.txt which contain categories that you want to detect.
- Now training
After change your classes_path in every file, then you can run train.py, the .pth files will be stored in /logs folder.
- Inference the results
Change the model_path and classes_path in yolo.py. Finally run predict.py to predict.
Main Code: https://github.com/bubbliiiing/yolov7-pytorch
Picture: Rongrong Lu, Haibo Sun, Shuangfei Fu, Feng Zhu, Yingming Hao. Point Cloud Registration Based Satellite Motion Parameter Identification Method[J]. Laser & Optoelectronics Progress, 2019, 56(14): 141503.