wrong result

[XJL705]

Hello, blogger, I would like to ask some questions about using your TEED model on my dataset. However, my dataset does not have ground truth data. How can I solve this issue?

Based on the above problem, I used the original images as the ground truth, but I am not sure why using your pre-trained weights resulted in relatively good results. However, when I load your trained weights and continue training my own experimental data, I do not get better results, as shown in the image below. Additionally, I have frozen all the convolution layers and only trained the last upsampling layers and fusion layers. How can I solve this issue?

[xavysp]

Hi, sorry for my late response, how many epochs did you train? Which one is the map trained with biped?