About image resize with intrinsics in config.

[QixingJiang]

Thank you for your contribution to the community through your work!
But I have some doubts about the code in config:
HFT/configs/pyva/pyva_swin_kd_simple_fpn_force_nuscenes.py

HFT/HFT/configs/pyva/pyva_swin_kd_simple_fpn_force_nuscenes.py

Lines 49 to 50 in c4d0671

	dict(type='LoadAnnotations', reduce_zero_label=False, with_calib=True, imdecode_backend='pyramid'),
	dict(type='Resize', img_scale=(1024, 1024), resize_gt=False, keep_ratio=False),

These configurations resize the images of nuScenes to 1024 and in the load function:

HFT/HFT/mmseg/datasets/pipelines/loading.py

Lines 202 to 207 in c4d0671

	if self.with_calib:
	token = osp.basename(filename).split('.')[0]
	intrinsics = torch.tensor(self.nuscenes[token])
	# scale calibration matrix to account for image downsampling
	intrinsics[0] *= 800 / results['img_shape'][1]
	intrinsics[1] *= 600 /results['img_shape'][0]

Here, the calib (intrinsics) is multiplied by a scaling factor, provided that the image is resized to a fixed size of 800600, not 10241024.
This hardcore caused some misalignment, did I miss something?
Thank you for answering my question.

[QixingJiang]

In fact, when the resize scale is 1024 * 1024, the camera parameters scaled to incorrect 600 * 800 in the current code will be better, and I don't know why ... Can you give me some clues?

[XYX1998]

你好，我想问下作者说的这个make_nuscenes_label.py生成的标签名字是token，请问训练的时候输入的是前视相机的图片吗