cvlab-unibo / mm2d3d Goto Github PK

hi @adricarda ,

thank you for your work.

I don't see the processing code for the A2D2 data set and semantic-kitti data set. Could you share the preprocess_a2d2.py and preprocess_semantic-kitti.py ?

Thanks!

hi @adricarda ,

thank you for your work.

I am so confused that why the experimental results shown in Tab. 1 are inconsistent with the results published in XMUDA?

The final mIoU of day->night UDA task is 50.0 in XMUDA, but you write 67.4 in your paper?

Hi, I rerun the code, but there is a gap between the experimental results and the results in the paper. The experimental results I got on the USA→ Singapore scenario: 2D-70.51, 3D-65.61, Avg-71.08. The results in the paper: 2D-71.7, 3D-66.8, Avg-72.4. Especially for the A2D2→ Sem.KITTI scenario, the resulting gap is large. Can you give me some suggestions?

Excuse me, a runtime error is reported when I run the script CUDA_VISIBLE_DEVICES=0,1 python experiments_USA_SING/rgbd_rgbxyz_sigmoid_for_rgb/run.py:
RuntimeError: mat1 and mat2 shapes cannot be multiplied (21584x1 and 3x1)

I run the lib/dataset/preprocess_nuscenes_lidarseg.py script to prepare the data.

I check and find that the data shape of data_batch["x"][1] is 21584x1 (located on line 46 of the experiments_USA_SING/rgbd_rgbxyz_sigmoid_for_rgb/3d_net/model.py file), but the mapping of the linear layer is 3 ->1:
mask_rgb = self.linear_rgb_mask(data_batch["x"][1])
self.linear_rgb_mask = nn.Linear(3, 1)

Also, can you clarify the version of the package (e.g. pytorch-lightning, hydra, etc.)?

Thanks for your reply and help!