search

yrcong / STTran

Spatial-Temporal Transformer for Dynamic Scene Graph Generation, ICCV2021
MIT License
181 stars 34 forks source link

Decoder_lin in Object Classifier #8

Closed Shengyu-Feng closed 2 years ago

Shengyu-Feng commented 2 years ago

Hi, I want to make sure for "sgdet" task, why is the decoder_lin not used during testing

Does it mean that the object prediction labels from the faserRCNN are used directly in "sgdet"? Then why is decoder_lin trained?

yrcong commented 2 years ago

Hi, the object classes are inferred by the FasterRCNN directly. I tried to train the custom classifier, hoping it can optimize the object classification. However, the custom classifier doesn't perform as expected. I think the reasons should be the bad quality of region proposals and overlapped object classes (e.g. bottle/cup).

You could remove the custom classifier for sgdet task by yourself :) This does not affect the result

Shengyu-Feng commented 2 years ago

Thanks for your reply!