search

nickgkan / butd_detr

Code for the ECCV22 paper "Bottom Up Top Down Detection Transformers for Language Grounding in Images and Point Clouds"
Other
74 stars 11 forks source link

Labels not updated for random_utt #18

Closed soham-joshi closed 1 year ago

soham-joshi commented 1 year ago

Hi @ayushjain1144 , While object detection is trained on ScanNet dataset, if there's a random utterance created why aren't the labels updated reference: line 730 in joint_det_dataset.py?

Thanks in advance!

nickgkan commented 1 year ago

Hi, this variable is used to fill the ret_dict['sem_cls_label'] field which is only used when evaluating object detection. Random detection prompts are used only during co-training with a referential grounding dataset. So ret_dict['sem_cls_label'] is not used in that case and thus _labels is also not used.

soham-joshi commented 1 year ago

Okay. Another query: is there a reason only a fixed set of classes (as in DC18 = ScannetDatasetConfig(num_classes=18) ) was selected for object detection?

nickgkan commented 1 year ago

The standard ScanNet object detection benchmark contains these 18 classes, see Group-Free/Votenet/etc, this setup is the same as theirs.

soham-joshi commented 1 year ago

Got it, Thanks for the response @nickgkan !