[Fix] Several Fixes for Multi-View 3D Grounding Baseline - Githubissues

OpenRobotLab / EmbodiedScan

[CVPR 2024] EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI

https://tai-wang.github.io/embodiedscan/

Apache License 2.0

395 stars 26 forks source link

[Fix] Several Fixes for Multi-View 3D Grounding Baseline #54

Closed mxh1999 closed 2 months ago

mxh1999 commented 2 months ago

Modification

Several fixes for multi-view 3D visual grounding baseline:

Augmentation: Fix the rotation and scaling behavior of EulerInstance3DBoxes (related to orientations), also affect MV-3DDet baseline
Data: Fix some data samples with inaccurate tokens_positive corresponding to the target in the language prompt
Model: Fix create_positive_maps for training with batch_size > 1
Enhancement: Support training with multiple grounding boxes in a single language prompt

All the fixes will result in a little performance improvement:

Method	AP@0.25	AP@0.5
Previous	35.15	12.75
Current	36.11	13.43