cshizhe / vil3dref

Official implementation of Language Conditioned Spatial Relation Reasoning for 3D Object Grounding (NeurIPS'22).
53 stars 5 forks source link

About detected object proposals on ScanRefer setting #1

Closed CurryYuan closed 1 year ago

CurryYuan commented 1 year ago

Hi @cshizhe ,

Thanks for your excellent work. Will you release the model at the setting of detected object proposals on ScanRefer?

Best

cshizhe commented 1 year ago

Hi, I uploaded the trained model using detected object proposals from PointGroup on ScanRefer, which achieves acc@iou0.25 0.479, acc@iou0.5 0.377. To be noted, the model using the groundtruth object proposals in training (in the original dropbox link in README) also achieves good generalization performance on detected object proposals with acc@iou0.25 0.466, acc@iou0.5 0.372.

CurryYuan commented 1 year ago

Hi @cshizhe ,

Thanks for sharing the trained model. I noticed that you provided the gt proposals. Could you also provide the detected proposals or share the processing code?

Best

cshizhe commented 1 year ago

You should first follow PointGroup to generate the segmentation results and then process the data into a similar format as the gt proposals. Some preprocessing codes are released here.