shiyoung77 / ovir-3d Goto Github PK

Hi, thank you so much for releasing your evaluation code for scanet200, could you release your evaluation code for ycb_video? Because the scanet200 dataset costs too much time to download.

hi, thank you so much for releasing your evaluation code!

I am running your evaluation code, but I seem to be receiving nan values. Any idea why this might be happening?

I am using a subset of the entire validation set. So perhaps that might be leading to the issue?

Hi, could you provide some guidance on how you generated the instance groundtruth labels for each point in the pointcloud in your scannet200 dataset? I am trying to do the same for the train dataset.

I notice that there are labels located in the aggregation.json file that comes with every scanned scene in scannet200. Can I simply take these for the GT instance labels?

Thank you so much ;))

Great project! I am trying to run your codebase on the scannet 200 dataset. I notice that there is a config.json file, but it does not seem to come with the scannet dataset. How should I generate this config.json file for scannet200 dataset?

Additionally, some files required for evaluating on scannet200 in readme seems to be missing. May I know when will they be released / how do I get access to them?

Thank you so much ;))

Hi, I notice that the current algorithm throws an error when it receives a frame of Detic prediction that is nothing.

Traceback (most recent call last): File "/mnt/src/proposed_fusion.py", line 633, in <module> main() File "/mnt/src/proposed_fusion.py", line 599, in main instance_pt_count, instance_features, instance_detections = instance_fusion( File "/mnt/src/proposed_fusion.py", line 283, in instance_fusion pred_masks = resolve_overlapping_masks(pred_masks, pred_scores, device=device) File "/mnt/src/proposed_fusion.py", line 233, in resolve_overlapping_masks indices = ((scores == torch.max(scores, dim=0, keepdim=True).values) & pred_masks).nonzero() IndexError: max(): Expected reduction dim 0 to have non-zero size.

How should I resolve this bug? I was thinking of simply skipping this iteration if mask dimension is 0, but I am not sure if it will mess up the rest of the algorithm. Thanks!

Hi @shiyoung77,

Thank you for your amazing work and the published code. I've encountered an issue while using your method on ScanNet200. Specifically, when visualizing all the generated instances in a scene (scene0011_00 in the validation split), the result appears as follows:

I've observed that many objects seem to be fragmented into multiple parts, which is considerably worse than in your paper. I'm wondering if there might be something missing in the implementation.

Thank you and looking forward to your reply.

Hi, just wanted to confirm my understanding, but it seems like the 3d instance segmentations having overlapping points? Is this meant to be the case? i.e each point in the point cloud can be assigned to multiple instance segmentations?

Hi, thank you so much for updating your codebase and fixing the issues!

I wanted to ask about the recent change that led to the drastic improvement.
Besides changing the depth threshold for occlusion detection, the filtering by size is changed from a fixed value to using the median. Does this mean that roughly half the instances are removed each filtering round?