Training config adjustment for large image

[oggyfaker]

Hi @Topdu , thanks for yours amazing repo !

• Currently, i got some issues related with the training and testing dataset with high resolution. My dataset shape are ranging from 3000 -> 5000 of Width and Height. Total dataset training i got is 5000 images.
• I used same config you defined here: config but my precision and recall are 0.8550 and 0.4718 respectively. Do you have any recommendation to change the config to get the better recall and precision (Image size , backbone para ...) ?
• And model usually misses the diagonal text and also include the triangle symbol in box prediction as image below. My training annotation is not including triangle symbol but i dont know the reason why model predict like this.

Sorry if any points makes you confusing, please let me know that. Hope to see yours answer soon !

[Topdu]

During training, you can try setting the input size to 1280x1280:
- EastRandomCropData:
size:
- 1280
- 1280
max_tries: 50
keep_ratio: true

During testing, try setting limit_side_len to 3200:
- DetResizeForTest:
limit_side_len: 3200
limit_type: max