search

fredzzhang / upt

[CVPR'22] Official PyTorch implementation for paper "Efficient Two-Stage Detection of Human–Object Interactions with a Novel Unary–Pairwise Transformer"
https://fredzzhang.com/unary-pairwise-transformers
BSD 3-Clause "New" or "Revised" License
150 stars 24 forks source link

suppress overconfident objects #85

Closed hutuo1213 closed 5 months ago

hutuo1213 commented 9 months ago

Hi,

We found that "suppress overconfident objects" works on HICO-DET but hardly works on V-COCO. Our work is in HICO-DET:

resnet50: λ=1.0 32.10mAP λ=1.9 33.44mAP λ=2.8 33.63mAP
resnet101: λ=1.0 32.48mAP λ=1.9 33.63mAP λ=2.8 33.79mAP

UPT in V-COCO:

resnet50: λ=1.0 58.9mAP λ=2.8 59.0mAP
resnet101: λ=1.0 60.7mAP λ=2.8 60.7mAP

Our existing view, this strategy works in HICO-DET and closes the gap between resnet50 and resnet101. Is it possible to find a unified view to explain the phenomena on HICO-DET and V-COCO?

fredzzhang commented 5 months ago

Hi @hutuo1213,

Sorry for the late reply! The choice of $\lambda$ was determined on the HICO-DET dataset through a simple hyper-parameter search. I simply used the same value for V-COCO. So it's not really optimised for V-COCO.

The use of $\lambda$ is more of a trick, so I did not investigate too much into it.

Fred.