yrcong / STTran

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

clean_class in ObjectClassifier #18

Closed jkli-aa closed 2 years ago

jkli-aa commented 2 years ago

Thanx for ur wonderful work. It seems that the cleanclass replaces the categories of those data that originally classified as class idx with their second highest confidence categories when it is sgdet mode. Could you explain why this operation performs here? Thanx again. Waiting for reply : )

yrcong commented 2 years ago

Hi! We processed like this due to the quality of the dataset. Some categories in the Action Genome are overlapped (somehow seriously), not only semantically but also spatially. For example in one frame there are two boxes respectively annotated as: class: book bbox:[138.94, 397.36,180.82,419.24] class:paper/notebook bbox:[138.93, 397.54,180.51,419.65]

To avoid the models suffering from this and make the result more convincing, we did this post-processing:)

taolunzu11 commented 1 year ago

你好,请问ObjectClassifier 中的 clean_class 的 class_idx 为何输入5,8,17呢