about underfitting

[lhj-git]

Thanks for your excellent work! I uniformly sampled 10 scenes from the training set and test set in ScanRefer. I found that the checkpoint you provided performs well on the test set（MIOU 0.40） but underperforms on the training set（MIOU 0.09） , showing signs of underfitting. What might be the reasons for this?
And I am wondering why the epoch of your training strategy is set to 1.

[zd11024]

Hi, thanks for your interest. After checking the data, I noticed that I did not provide camera parameters for the training set. So, how do you calculate the mIoU on the training set? For the validation set, we first transform the bounding box in the camera's coordinate to the world coordinate and then compare the IoU with the ground truth bbox.

[lhj-git]

Assuming my current outputs are frame F and 3d_box B:

In the Embodiscan dataset, locate the corresponding camera extrinsic parameters based on the image path associated with the frame.

Obtain the camera extrinsic parameters by calculating axis_align_matrix * cam2global.

Transform B to the GT (ground truth) using the camera extrinsic parameters.

Apply the same transformation to the GT frame F_GT and 3d_box B_GT.

Calculate the IOU.

[zd11024]

I have tested the visual grounding task on the training set, and the results are as expected. I utilized the convert_scanrefer.py script to generate the training annotations, with necessary camera parameters.