Space object detection

Based on SMA and YOLO V7 and Blender Simulation

Simulating data with Blender

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.

Annotating dataset with ISAT_with_segment_anything

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.

Training and Inferencing with YOLO V7

1. 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

2. 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.

3. 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.

4. Inference the results

Change the model_path and classes_path in yolo.py. Finally run predict.py to predict.

Reference

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.