search

zlccccc / 3DVL_Codebase

[CVPR2022 Oral] 3DJCG: A Unified Framework for Joint Dense Captioning and Visual Grounding on 3D Point Clouds
Other
51 stars 5 forks source link

Unable to train with RGB inputs #7

Closed ch3cook-fdu closed 1 year ago

ch3cook-fdu commented 1 year ago

File "/root/3DJCG/models/proposal_module/relation_module.py", line 111, in forward obj_embedding = self.obj_embeddingi

input dimension error

ch3cook-fdu commented 1 year ago

The implementation in models/proposal_module/relation_module.py, line 95 - 97 is incorrect for RGB inputs, since PLC_COLOR is stored in data_dict["point_clouds"][..., 3:6]

# multiview/rgb feature embedding
if self.use_obj_embedding:
    obj_feat = data_dict["point_clouds"][..., 6:6 + 128].permute(0, 2, 1)
zlccccc commented 1 year ago

Fixed now.