Issue with detecting borderless headers

[KaranBhuva22]

Hello @AlexeyAB @WongKinYiu, I tried to fine-tune yolov7 using the yolov7-w6.pt pretrained weights. Overall, the model performs well, although it struggles to detect some borderless table headers despite having borderless table images in the training set.

The observed pattern is that the model accurately predicts some borderless headers within the table but fails to identify several intermediate headers. For instance, if there are 10 headers in the table, the model may successfully predict 4 or 5 headers while overlooking the remaining 5 or 6 headers.

This problem only occurs with borderless headers, model detects bordered headers perfectly.

Model Configurations : Training resolution : 1280 Batch size : 16 Total Images in training : 16K

The following are examples of borderless tables in which the model fails to detect headers.

What are the possible ways for improvements ?

[dsbyprateekg]

@KaranBhuva22 what is the --conf-thres you used in the above images?

[KaranBhuva22]

@dsbyprateekg --conf-thresh is 0.7

[dsbyprateekg]

@dsbyprateekg --conf-thresh is 0.7

Check with a lower one like 0.5 or default one 0.25.