jackroos / VL-BERT

Code for ICLR 2020 paper "VL-BERT: Pre-training of Generic Visual-Linguistic Representations".
MIT License
738 stars 110 forks source link

Improper RefCOCO evaluation #83

Open st2yang opened 3 years ago

st2yang commented 3 years ago

When comparing MAttNet in Table 3, the target classification accuracy should be used. While computing IOU to approximate the classification accuracy, a small threshold (https://github.com/jackroos/VL-BERT/blob/master/refcoco/function/test.py#L17) was used, which could count false positive as correct.