search

BruceW91 / CVSE

The official source code for the paper Consensus-Aware Visual-Semantic Embedding for Image-Text Matching (ECCV 2020)
172 stars 19 forks source link

about evaluation.py #10

Closed gedaye11 closed 3 years ago

gedaye11 commented 3 years ago

Hi , why you used "img_emb_inst = img_emb[2] cap_emb_inst = cap_emb[2] " in the line 96,97 of evaluation.py？？？ In my opinion, the instance-level features is img_emb[0] and cap_emb[0] ???????

because in Model.py, the code is "emb_v = torch.stack((instance_emb_v, consensus_emb_v, fused_emb_v), dim=0)".

this confused me a lot. Hope your reply.

BruceW91 commented 3 years ago

Hi, this configuration is setted by emperically finding it brought about slight performance improvement when the instance-level features are replaced by the fused features in my subsequent experiments.