Some annotations have overlapping point indices

[chrockey]

Hi @AlexDelitzas, thank you for releasing the SceneFun3D dataset — it’s been very helpful for my research!

While exploring the annotations, I noticed that some of them contain overlapping point indices, i.e., the same point is assigned to multiple annotations as shown below:

➜  /workspace git:(main) ✗ python -m src.data.scenefun3d.check_annotation --data_dir /datasets/SceneFun3D --metadata_dir src/data/metadata/split_files                                                                                                                  [0/0]
                                                                   
Starting annotation validation...
train: 200 scenes             
val: 30 scenes                                                                                                                        

  Checking 468476 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100%
                                                                                                                                      
Annotation overlap statistics:
                                                                   
train split:                                                                                                                          

  422523 (total points: 2279):
    Annotations 0ef692ea-7702-4392-9013-66e7ca19d865 and 20a99835-9b8f-4369-9484-39945584409c share 1 points  

  464236 (total points: 3938): 
    Annotations 5359be1b-5bec-42e9-b270-c76bd861fec1 and bd94757d-885c-4f60-b341-384a2f6b804f share 1 points

  466176 (total points: 5395):
    Annotations 76d8bbd9-7504-4887-828b-8aa2a32032fd and 94f8a440-a449-44cd-8df1-dc9aacea1207 share 2 points
                                                                   
  468071 (total points: 1221):                                                                                                        
    Annotations 3d464909-1e4d-4d2a-bd04-594b6373fb93 and d81654ec-1e21-4648-873f-cf4e69f8c6ec share 257 points
                                                                   
  437448 (total points: 16135):                                                                                                                                                                                                                                              
    Annotations 681eb03e-62f7-4078-9a7a-f15ee6e953db and b5fa1b8f-6ec4-4ac6-b27f-cebc48bfe26f share 2 points

val split:

  467115 (total points: 13519):
    Annotations 9ab26fb5-ae61-4977-b743-e081ae50c40b and ee8a1b91-c6e8-4739-a4af-120ef76858e9 share 1 points

Found 6 scenes with overlapping annotations

I wanted to ask if this is expected behavior (e.g., for overlapping regions or multi-label annotations), or if this might be an issue in the dataset.

In particular, in the case of visit_id: 468071, where 257 points are shared between two annotations, I visualized the scene and found that the annotation labeled as “hook_turn” seems to correspond to a door handle, but also includes a nearby light switch. You can see this in the screenshot below:

annot_id: 3d464909-1e4d-4d2a-bd04-594b6373fb93

annot_id: d81654ec-1e21-4648-873f-cf4e69f8c6ec

Any clarification on this would be greatly appreciated. Thanks again for your great work!