Firstly, download PubTables-1M dataset and put it into ./PubTables-1M.
Train and val:
# Train Detection
python train.py --batch-size -1 --data data/custom-detection.yaml --weights yolov5m.pt --img 640 --project PubTables-1M-YOLO --name yolov5m-custom-detection --hyp data/hyps/hyp.scratch-med.yaml --epochs 15 --device 0
# Val Detection
python val.py --batch-size 256 --data data/custom-detection.yaml --weights PubTables-1M-YOLO/yolov5m-custom-detection/weights/best.pt --project PubTables-1M-YOLO --name yolov5m-custom-detection-val --img 640 --task test --device 0
# Train Structure
python train.py --batch-size -1 --data data/custom-structure.yaml --weights yolov5m.pt --img 640 --project PubTables-1M-YOLO --name yolov5m-custom-structure --hyp data/hyps/hyp.scratch-med.yaml --epochs 15 --device 0
# Val Structure mAP
python val.py --batch-size 256 --data data/custom-structure.yaml --weights PubTables-1M-YOLO/yolov5m-custom-structure/weights/best.pt --img 640 --project PubTables-1M-YOLO --name yolov5m-custom-structure-val --task test --device 0
# Val Structure GriTS
python val_structure_grits.py --batch-size 256 --data data/custom-structure.yaml --weights PubTables-1M-YOLO/yolov5m-custom-structure/weights/best.pt --img 640 --project PubTables-1M-YOLO --name yolov5m-custom-structure-val --task test --conf-thres 0.25 --iou-thres 0.45 --device 0Table Detection:
| Model | Schedule | AP50 | AP | Precision | Recall |
|---|---|---|---|---|---|
| YOLOv5m | 15 Epochs | 0.995 | 0.988 | 0.998 | 0.999 |
Table Structure Recognition:
| Model | Schedule | AP50 | AP | Precision | Recall | GriTSTop | GriTSCon | GriTSLoc | AccCon |
|---|---|---|---|---|---|---|---|---|---|
| YOLOv5m | 15 Epochs | 0.973 | 0.912 | 0.908 | 0.950 | 0.9826 | 0.9828 | 0.9751 | 0.8128 |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent