several preliminary experiments for videl tracking task.
- Linux (Ubuntu 18)
- Python ≥ 3.6
- Opencv-contrib-python > 4.1.0 recommended
Take CSRT algorithm as an example, run tracking.py with:
python tracking.py --video tracking.mp4 --algorithm CSRT
Dataset is from deep-dental-image.
All the preliminary experiments are based on pretrained models and opensource packages.
-
Cuda 10.1 guidance 1 guidance 2 guidance3.
-
Facebook detectron 2.
-
Git shape_to_coco to path /detectron2/.
tips:
- Driver, cuda and pytorch should be matched perfectly.
- copy all the .py files to detectron/, rename tooth dataset and put it to detectron/dataset/.
- generate train and val dataset from the original UFBA_UESC_DENTAL_IMAGES_DEEP dataset.
cd detectron2
python division_train_val.py
- generate annotation .json files for train and val datasets seperately. waiting for 15 minutes to generate two json files.
python shape_to_coco.py
- run tooth_train.py to train and evaluate the detection performance.
Mask rcnn is not suitable for tooth semantic segmentation since the annotations are coarse and do not support instance segmentation. The network are confused by the shape of teeth. If you are interested in tooth semantic segmentation in panoramic X-ray iamges and CBCT images, please click [here] (https://github.com/liangjiubujiu/stage_1) or find more information in the author's page.
Simply use the MASKRCNN pretrained model and guideance. Predited categories and segmentation annotations are not ideal.
- Simply compare the results in this pretrained model and in published paper. They looks very different.
** This pretrained model prefer category No.8 in category No. 0~9.
- Ps: All therepository is made by the detailed markdown guidance.