search

xingyizhou / CenterTrack

Simultaneous object detection and tracking using center points.
MIT License
2.38k stars 526 forks source link

Question about train own dataset #66

Open LiXm1002 opened 4 years ago

LiXm1002 commented 4 years ago

Thanks for sharing! I want train a tracking model on my own dataset with coco_tracking.sh in experiments folder. But my dataset has only one category. Where should I modify the coco 80 category dict to my category dict and use it to train a model?

xingyizhou commented 4 years ago

Hi, You will need to prepare a json file in COCO format. See src/tools/convert_xxx_to_coco.py for references. After that, you don't need to write any code. Just add --dataset custom --custom_dataset_ann_path /path/to/your/annotation.json --custom_dataset_img_path /path/to/your/image/folder/ --input_h XX --input_w XXX --num_classes 1 to your training command.

LiXm1002 commented 4 years ago

Thanks for reply. I have used Lableme to annotate some pictures in coco format to train a instance segmentation model. Can I directly use this data to train CenterTrack? Or still use convert_xxx_to_coco.py to convert?

Maeve-D commented 4 years ago

@LiXm1002 Have you solved the problem？