Danila Rukhovich
Danila Rukhovich
I think, for ScanNet it is possible to use encoder as we initialize queries with all superpoints. However for S3DIS i don't have good ideas how to use only encoder,...
Don't remember such experiments. I think for 20 images it should not be much difference, but 6 is too small.
Actually this idea requires more efforts compared to single image + single depth on SUN RGB-D. You need somehow to project image features to points then aggregate them. We never...
I think, we only support point cloud visualization it his repo.
You can run visualization command from our readme and than open the results in meshlab.
What do you mean? If there are many extra low confidence predictions you can increase `score_thr` in config. If you want points to be colored and the boxes to be...
Can you share the full config `.py` file? Also to be sure, train and validation is ok, but test is not ok?
Ah, this `pcd_demo.py` thing doesn't work i think. You need to debug it a little bit, to be sure that everything is exactly like in `test.py` script. One of the...
Btw i'm a little confused about rotation. As you use `ScanNetDataset` and `n_reg_outs=6` in your config, we don't even predict rotation in this case. So all roations are zeros and...
Hard to say what is wrong with your dataset, as you don't give much details. I recommend to tune `voxel_size`, `n_points`, and `samples_per_gpu` in the config file.