I follow your instructions and run your pretrained model. But the eval result is iou0.25: 0.3751 iou0.5: 0.3056 I do not know why it is different form yours. Thank you very much!

Thank you for this amazing work.

When I run "python scripts/train.py --log_dir instancerefer", I encountered a problem:

File "/home/nicole/InstanceRefer/lib/dataset.py", line 184, in getitem class_ind = [DC.nyu40id2class[int(x)] for x in instance_bboxes[:num_bbox, -2]] File "/home/nicole/InstanceRefer/lib/dataset.py", line 184, in class_ind = [DC.nyu40id2class[int(x)] for x in instance_bboxes[:num_bbox, -2]] KeyError: 0

the x is 0.0 but DC.nyu40id2class does not contain 0. Should I modify the code as "DC.nyu40id2class[int(x)+1]"?

In Step 5. Pre-processed instance labels, and new data should be generated in data/scannet/pointgroup_data/

Since train and test data of scannet are placed at different positions, the commands should be:

cd data/scannet/
python prepare_data.py --split train --scannet_path REPO_DIR/data/scannet/scans --pointgroupinst_path [YOUR_PATH]
python prepare_data.py --split test  --scannet_path REPO_DIR/data/scannet/scans_test  --pointgroupinst_path [YOUR_PATH]

Hi there,

the original _vh_clean.aggregation.json is for the high-res mesh files - I've corrected it to .aggregation.json which matches the low-res files you use in this repo.

(plz check ScanNet README)

Best Dave

I am currently having a look at the performance of your architecture. In the paper you state that the [email protected] is 37.05 using only xyz coordinates which is way higher than the [email protected] of 30.70 using additional color information. Is this just a typo in the paper or is the additional rgb information really damaging your performance? If this is a typo, could you please provide me with the correct [email protected] results?

Thank you very much!

Hi, CurryYuan: I have met some problems as below: Preparing data... train on 36665 samples and val on 9508 samples loading data... loading data... Initializing... model params 6857788 Start training... Epoch 1 starting... Traceback (most recent call last): File "/remote-home/source/46/guopeng/scripts/train.py", line 221, in train(CONF) File "/remote-home/source/46/guopeng/scripts/train.py", line 211, in train solver(args.epoch, args.verbose) File "/remote-home/source/46/guopeng/lib/solver.py", line 156, in call self._feed(self.dataloader["train"], "train", epoch_id) File "/remote-home/source/46/guopeng/lib/solver.py", line 248, in _feed for idx, data_dict in enumerate(dataloader): File "/usr/local/miniconda3/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 363, in next data = self._next_data() File "/usr/local/miniconda3/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 989, in _next_data return self._process_data(data) File "/usr/local/miniconda3/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 1014, in _process_data data.reraise() File "/usr/local/miniconda3/lib/python3.6/site-packages/torch/_utils.py", line 395, in reraise raise self.exc_type(msg) RuntimeError: Caught RuntimeError in DataLoader worker process 0. Original Traceback (most recent call last): File "/usr/local/miniconda3/lib/python3.6/site-packages/torch/utils/data/_utils/worker.py", line 185, in _worker_loop data = fetcher.fetch(index) File "/usr/local/miniconda3/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py", line 47, in fetch return self.collate_fn(data) File "/remote-home/source/46/guopeng/lib/dataset.py", line 464, in collate_fn outputs = sparse_collate_fn(inputs) File "/usr/local/miniconda3/lib/python3.6/site-packages/torchsparse-1.1.0-py3.6-linux-x86_64.egg/torchsparse/utils/helpers.py", line 209, in sparse_collate_fn dim=0) RuntimeError: zero-dimensional tensor (at position 0) cannot be concatenated I think maybe you know the Bug.

Hi, thanks for your amazing work.

We notice that in your newly released version, you use co-attention from MCAN to achieve a much higher performance, I wonder that will this part of code be released?

Thanks!