Mmdetection model fine tuning

Use FasterRCNN Training model , Because what we do is taxi recognition in traffic scenes , I marked some data , To enhance the effect , First in the dataset BDD100K Training , Then fine tune your data set .
Use ·faster_rcnn_r50_fpn_1x Training

BDD100K Dataset download address ：https://doc.bdd100k.com/download.html
Model download address ：https://github.com/SysCV/bdd100k-models
By looking at the document , The trained model provided by the government is similar to MMDetection A match , Don't train yourself , You can download it directly .
faster_rcnn_r50_fpn_1x The corresponding model download address is ：
https://dl.cv.ethz.ch/bdd100k/det/models/faster_rcnn_r50_fpn_1x_det_bdd100k.pth

modify MMDetection Code , The modification of data sets and common problems can be referred to

MMDetection Use and related problems

Compared with direct training , Fine tuning only requires modification configs/base/default_runtime.py in load_from = None, take None Change the path to download the model , Or directly copy the above download address

load_from = "work_dirs/faster_rcnn_r50_fpn_1x_det_bdd100k.pth"

Other operations are the same as training your own dataset .

MMDetection Fine tune document reference docs/tutorials/finetune.md

https://github.com/open-mmlab/mmdetection/blob/master/docs/tutorials/finetune.md