Sharing some scene issues with the dataset

[yycho0108]

Dear 3D-FRONT staff,

Hi, this is Yoonyoung (Jamie) Cho, researcher at the KAIST Intelligent Mobile Manipulations Lab.
Again, thank you for sharing this dataset - the 3D-FRONT dataset will be an invaluable asset in our research.
Just for context: I have previously contacted the team via email, but I figured I should move the conversation here for public discussion.

I have noticed some issues throughout my interaction with the dataset. For public reference, I have validated that the checksum of my local copy matches that of the source. The following command describes the checksum at the time of writing (2021-03-03):

$ md5sum 3D-FUTURE-MODEL.zip
cf20ec68ce7005d5491d2f207a7730e6 3D-FUTURE-model.zip

$ md5sum 3D-FRONT.zip
d8d7521283365c56e5da3ce247fe8006 3D-FRONT.zip

Now that we have established that the errors in the dataset are not due to an error during the download process, I'd like to enumerate some issues that I have noticed over time.

• Some models that contain the Shadow attribute are broken. An example can be found 3D-FUTURE-model/3D-FUTURE-model/6914bc9b-d7f3-3d85-9fdf-a8e30b54ed35/raw_model.obj; the scale is severely off, as it can be seen below. An example of a scene that contains this model is 0ec7ce97-e93d-4842-9c89-b947136bb393.json.
  
  (In the above scene, the scale of the dining chair is severely off, and it's towering above the whole house.)
  Note that this is not an error that can be caught with the default json2obj.py which appears to silently ignore this issue. The real problem is that when loading this scene in programmatically through other interfaces such as pybullet, this scene will cause issues. My current guess for the root cause here is that, in this case, the .obj file is parsed as a scene - which possibly violates some hardcoded assumptions in the scene processing pipeline.

• Some of the texture images appear to be broken. An example can be seen in 3D-FUTURE-model/6934dea0-1d66-49c4-82c6-4d54d41f9707/texture.png. (libpng error: Read Error) and failed to be loaded in tools such as meshlab. Here's what the texture looks like when I visualize it with eog:
  

• One of the models have an error: 3DFRONT/3D-FUTURE-model/3D-FUTURE-model/e73ff703-adb1-4d3a-993d-60f6d148bec4/raw_model.obj an example of a scene that includes this model is 3D-FRONT/99e87673-0512-41c8-b35e-fd1ea39d1f0c.json.

  • The offending line is (raw_model.obj#L109726): f 6970/. Removing this line fixes the issue but I'm not sure what the initial intent was.

• Some of the scenes have an apparently empty mesh, such as 3DFRONT/3D-FRONT/6ce404bd-9c24-4a3b-869f-5217310e3741.json. The relevant snippet has been copied below (after formatting):

        {
            "aid": [],
            "jid": "",
            "uid": "49261587189179756/0",
            "xyz": [],
            "normal": [],
            "uv": [],
            "faces": [],
            "material": "53951587189179627/05395",
            "type": "WallOuter",
            "constructid": "4832",
            "instanceid": "4926"
        }

• Not sure if this is intentional, but many of the scenes are missing models. This seems to be particularly the case for certain classes of objects (door/window). For instance, when I try to load an example scene 3DFRONT/3D-FRONT/f24f4eab-5da2-4308-9145-fad49fc3916b.json, the following shapes are not found:

8799/model
8805/model
12697/model
159369944712376608/model
8806/model
18408/model
20553/model
20553/model
20553/model
20553/model
20553/model
20553/model
20553/model
20553/model
20553/model
20553/model
21419/model
159369944712999808/model
21433/model

• Finally, because all scenes are essentially missing all door-like objects, every floor has a gap where a door should have been placed. See the below picture for reference (shown without texture for clarity):
  
  This essentially means that nearly every scene in 3DFRONT has a discontinuity in the scene floor.

I hope the enumeration of the above set of issues was helpful! Let me know if there is anything I can do to help. Again, I'd like to stress how much I appreciate that you have shared the dataset with us. I think the dataset really has strong potential, and after some of the above issues have been addressed the dataset would be a lot easier to use for some of us.

Thank you,

Yoonyoung (Jamie) Cho