search

yangyangyang127 / PointCLIP_V2

[ICCV 2023] PointCLIP V2: Prompting CLIP and GPT for Powerful 3D Open-world Learning
MIT License
225 stars 13 forks source link

shape of feat_store and test_feat on task zeroshot_seg #23

Open kingskymoon opened 10 months ago

kingskymoon commented 10 months ago

Hello, thank you for your excellent open-source work, I hava a question about feat_store and test_feat on task zeroshot_seg.

First, feat_store is generated by torch.save(feat_store, os.path.join(save_path, "{}_features.pt".format(mode))) torch.Size([14, 10, 512, 14, 14]) 14 is batch,10 is num of views, 512, is channel of feature dim, and image width is 14*14

Then, test_feat = test_feat.reshape(-1, 10, 196, 512)

Why is it not necessary to move the features to the last dimension?

test_feat = test_feat.permute(0,1,3,4,2)
test_feat = test_feat.reshape(-1, 10, 196, 512)