Pretrained Weights
Model Zoo
| Source Dataset | Model | Epoch | File | Size |
|---|---|---|---|---|
| CTW-1500 | textgraph_vgg | 2964 | textgraph_vgg_2964.pth |
~147 MB |
Details
- CTW-1500 is a large-scale dataset for curved/scene text detection with 1,000 training images and 500 test images.
- This checkpoint was trained for 2,964 epochs on CTW-1500, then used as initialization for LATeD fine-tuning.
- The pretrained model captures general text detection features: TR/TCL segmentation maps, sin/cos geometry prediction, and bounding box regression.
- Transferring CTW-1500 → LATeD accelerates convergence on the low-light domain.
Usage
In configs/train_LATeD.yaml:
model:
pretrained_weight: ./checkpoints/pretrained/textgraph_vgg_2964.pth
Download
Place the checkpoint file in this directory. The file should be named:
checkpoints/pretrained/textgraph_vgg_2964.pth