Ball Video Augmentation Training
Screenshot
Capture screenshots in pot player and store them in a unified location.
Retrieve Source Video Screenshots
/home/liying_lab/chenxinfeng/ml-project/LILAB-py/lilab/outlier_refine/process_ball_dzy.sh
Labeling
Use labelme for labeling. This is an example of a labelme annotated json file.
{
"version": "4.2.5",
"flags": {},
"shapes": [
{
"label": "ball",
"points": [
[
226.4642857142857,
410.74999999999994
]
],
"group_id": null,
"shape_type": "point",
"flags": {}
}
],
"imagePath": "2024-02-27_21-59-28ball_bob_room2_pannel8_002675.jpg",
"imageData": null,
"imageHeight": 800,
"imageWidth": 1280
}
Important Warning
Please use labelme's point
type, not rectangle
type, as the point type can more accurately label the ball's position. The label should be named ball
. Please ensure the correct label name is used during labeling.
Rename the Data Folder and Place it in the Specified Location
/home/liying_lab/chenxinfeng/DATA/mmpose/data/ball
labelme Format Conversion
python -m lilab.cvutils.labelme_to_cocokeypoints_ball $LABEL_IMG_DIR
Training
/home/liying_lab/chenxinfeng/DATA/mmpose/res50_coco_ball_512x320_cam9.py Modify the code to add the new data folder and json file address to the new dict.
cd /home/liying_lab/chenxinfeng/DATA/mmpose
python tools/train.py res50_coco_ball_512x320_cam9.py
Model Acceleration
/home/liying_lab/chenxinfeng/ml-project/LILAB-py/lilab/mmpose_dev/convert_model.sh Modify the corresponding model filename to the variable.
Confirmation
Check if there are new json files, etc. in /home/liying_lab/chenxinfeng/DATA/mmpose/work_dirs and the update time of latest.*
Re-run the Ball Positioning Code from Scratch
/home/liying_lab/chenxinfeng/ml-project/LILAB-py/lilab/multiview_scripts_dev/calibration.sh