fudan-zvg / DeepInteraction

[NeurIPS 2022] DeepInteraction: 3D Object Detection via Modality Interaction
MIT License
221 stars 16 forks source link

question about the I2P block #33

Open HydrogenWasser opened 1 year ago

HydrogenWasser commented 1 year ago

In your paper, about the part of attention output of I2P block, it said the query should be image feature.

image

but in your code, encoder_utils, line 284,

Q = lidar_feat[b,:,voxel_coor[:,2].long(),voxel_coor[:,3].long()].t().unsqueeze(1)

it seems the query is lidar feature but not image feature