search

skanti / SceneCAD

[ECCV 2020] SceneCAD: Predicting Object Alignments and Layouts in RGB-D Scans
76 stars 5 forks source link

Evaluation table in paper (table 1) conflicts with previous studies #6

Open jeonghyunkeem opened 3 years ago

jeonghyunkeem commented 3 years ago

It seems the order of object category in table 1 from the paper is different from the previous studies(Scan2CAD, E2E) while values are matched perfectly.

SceneCAD: bathtub bookshelf cabinet chair display other sofa table trashbin S2C, E2E: bath bookshelf cabinet chair display sofa table trash bin other

Is it just a mistake or is there any intention about it?

Also, how did you train an object detection network using 3D-SIS with different object labels? I guess the display and trashbin category is considered the same as otherfurniture in scannetv2. Can you provide more details about it?

skanti commented 3 years ago

Hi @jeonghyunkeem, thanks for pointing that out! It's a mistake - it should be the same class with the same values. We will try to correct it.

Regarding your 2nd question: We did not use 3D-SIS as detector. Generally in CAD alignment tasks you do not need to predict the class labels since you are going to align a CAD model to the scan and thus can simply assume the scan object has the same class as the used CAD model. Class labels in our work are used mainly for evaluation. Hope that helps!